YOLO-World in Python: Detect Objects with Text Prompts

Can an object detector look for categories you describe in words? In this Python tutorial we use a pretrained YOLO-World model and an image, choose text labels, and inspect the bounding boxes it returns. We then change the label list and run the model again.

Our example uses a synthetic office image from my course and a local Windows CPU run. With one list the model returned four predictions; with an expanded list it returned nine predictions. Those counts are model outputs, not verified counts of real objects or proof of accuracy.

Watch the YOLO-World tutorial

Watch the full Python tutorial or see the short demonstration.

Actual focused-list detections from the recorded run. This picture is a synthetic teaching example.

What does “no training needed” mean?

YOLO-World is an open-vocabulary object detector: it can accept text class names before prediction. Here we load pretrained YOLO-World and CLIP components and do not train them again. They were trained earlier; “no additional training” does not mean that the model learned from nothing. It also does not guarantee detection of every item or phrase. See the Ultralytics YOLO-World documentation and the original research paper.

What you need

The demonstrated run used Windows 11, Python 3.13.9, Ultralytics 8.4.116, PyTorch 2.13.0+cpu, OpenCV 4.14.0 and NumPy 2.5.1. It used the local YOLOv8s-World checkpoint and the OpenAI CLIP ViT-B/32 text encoder. The installation commands pin Ultralytics and the CLIP fork, not every transitive dependency. Other software versions or devices can change the results. Git is needed for the command that installs the fork from its repository.

From a Windows Command Prompt, create a project folder, then create and install into its virtual environment:

py -3.13 -m venv .venv
.venv\Scripts\python.exe -m pip install ultralytics==8.4.116
.venv\Scripts\python.exe -m pip install "git+https://github.com/ultralytics/CLIP.git@488e81a6711eea7346872b46ea928b367da8889d"

The tested course package includes an already built source-bound wheel for the CLIP fork. The Git-based installation above is an alternative way to get the identified source; this exact fresh online installation was not part of the offline replay. Installing packages and obtaining the files requires a network connection unless you have them available locally.

Prepare the local image and weights

Save the two scripts below next to a permitted image called image.png. The shown image is supplied with lesson 47 of my Computer Vision course. For your own image, the detections and numbers will differ. Download the official Ultralytics YOLOv8s-World checkpoint and the OpenAI CLIP ViT-B/32 checkpoint separately. The latter URL appears in the OpenAI CLIP model registry. Place them as follows:

yolo-world-demo/
├── setup_local.py
├── world_demo.py
├── image.png
├── yolov8s-world.pt
└── weights/
    └── clip/
        └── ViT-B-32.pt

The tested detector checkpoint SHA-256 is 095f5266bb9b654bd5ad9e21e9cdeda78e0f2c8460f5d652eaf04bab7ee251cf; the CLIP file SHA-256 is 40d365715913c9da98579312b702a82c18be219cc2a73407c4526f58eba950af. A hash identifies bytes; it does not itself establish usage rights. Model files are not hosted by this article.

Configure a project-local weights folder

The text encoder must find its local CLIP checkpoint when the class names are set. Run the following small setup script once, as a separate process. It places Ultralytics settings in this project’s config folder and sets its weights directory to this project’s weights folder. This avoids editing your user-wide Ultralytics settings for the demonstration.

Save as setup_local.py:

# SPDX-License-Identifier: AGPL-3.0-only
# Copyright (c) 2026 ARTSCI - Author of the AI/ML/CV course materials.
"""Store an isolated Ultralytics weights-directory setting for this example."""
from pathlib import Path
import os

ROOT = Path(__file__).resolve().parent
config_dir = ROOT / 'config'
config_dir.mkdir(exist_ok=True)
os.environ['YOLO_CONFIG_DIR'] = str(config_dir)
import cv2  # Import first in the demonstrated Windows environment.
from ultralytics import settings

settings.update({'weights_dir': str(ROOT / 'weights'), 'sync': False})
print('Project weights directory:', ROOT / 'weights')
print('Run world_demo.py as a separate process after this one exits.')

Run it from the project directory:

.venv\Scripts\python.exe setup_local.py

Then start the following script as a new process so the weights-directory setting is read during initialization. If you move the project folder later, run the setup script again before prediction.

Complete runnable example

Save the following as world_demo.py:

# SPDX-License-Identifier: AGPL-3.0-only
# Copyright (c) 2026 ARTSCI - Author of the AI/ML/CV course materials.
"""Compare two text vocabularies with a local pretrained YOLO-World model."""
from pathlib import Path
import os

ROOT = Path(__file__).resolve().parent
# Keep Ultralytics settings in this project, not in your user-wide profile.
config_dir = ROOT / 'config'
config_dir.mkdir(exist_ok=True)
os.environ['YOLO_CONFIG_DIR'] = str(config_dir)
os.environ['YOLO_OFFLINE'] = 'true'

import cv2  # Import first in the demonstrated Windows environment.
import torch
from ultralytics import YOLOWorld
from ultralytics.utils import WEIGHTS_DIR


def main():
    image = ROOT / 'image.png'
    model_file = ROOT / 'yolov8s-world.pt'
    clip_file = ROOT / 'weights' / 'clip' / 'ViT-B-32.pt'
    if Path(WEIGHTS_DIR).resolve() != (ROOT / 'weights').resolve():
        raise RuntimeError('Configure this project weights directory first; see the setup command.')
    for path in (image, model_file, clip_file):
        if not path.is_file():
            raise FileNotFoundError(f'Required local file: {path}')

    torch.set_num_threads(4)
    model = YOLOWorld(str(model_file))
    groups = {
        'focused': ['person', 'backpack', 'laptop', 'cup'],
        'broad': ['person', 'backpack', 'laptop', 'cup', 'chair', 'desk'],
        'single_cup': ['cup'],
        'single_coffee_cup': ['coffee cup'],
    }
    output = ROOT / 'output'
    output.mkdir(exist_ok=True)
    for name, classes in groups.items():
        model.set_classes(classes)
        result = model.predict(
            source=str(image), conf=0.25, iou=0.70, imgsz=640, device='cpu',
            half=False, rect=True, augment=False, agnostic_nms=False,
            max_det=300, verbose=False,
        )[0]
        print(f'\n{name}: {classes} -> {len(result.boxes)} predictions')
        for box in result.boxes:
            label = result.names[int(box.cls.item())]
            score = float(box.conf.item())
            coords = [round(float(v), 1) for v in box.xyxy[0].tolist()]
            print(f'  {label}: score={score:.3f} xyxy={coords}')
        annotated = result.plot()
        if not cv2.imwrite(str(output / f'{name}.png'), annotated):
            raise OSError(f'Could not save output for {name}')
    print(f'\nSaved annotated images in {output}')


if __name__ == '__main__':
    main()

From the project directory, run:

.venv\Scripts\python.exe world_demo.py

For each vocabulary, model.set_classes(classes) configures the class names with the CLIP text encoder. model.predict(...)[0] then runs a new prediction on the image; the subsequent loop reads its boxes. The code saves an annotated image per vocabulary to output/. The illustrations on this page are the actual saved outputs from the demonstrated execution, using Ultralytics’ plotting style.

All four runs use the same input image and checkpoint, conf=0.25, iou=0.70, imgsz=640 and CPU. The two standalone prompts are separate one-class vocabularies; the focused and broad lists are separate runs too. Changing the list can change predictions and scores for shared classes, so the broad result should not be treated as simply appending new boxes to an unchanged prior result. Confidence is a model score, not a calibrated probability that a particular box is correct.

Read the result: focused versus broad

The focused list asks for person, backpack, laptop and cup. In the recorded execution it returned four predictions. The broad list retains those words and adds chair and desk; it returned nine predictions. The change in count does not establish that every additional prediction is correct.

Actual broad-list output. Inspect overlaps and questionable boxes rather than equating nine predictions with nine objects.

For example, the expanded run includes a desk box scoring approximately 0.286 over a region that appears to contain a chair. Another nearby chair prediction overlaps the region. This is a reason to inspect the image and outputs, not evidence that every overlap is a duplicate. Changing the vocabulary also changed the laptop score from about 0.924 to 0.909 in this particular run.

Does wording matter?

We also tried two separate one-class runs on the same image. With ["cup"] the model returned one prediction with score approximately 0.917. With ["coffee cup"] it returned one prediction around the same visible item with score approximately 0.830. These two scores and boxes are observations for this image and these prompts; they do not prove that either phrase is generally better or measure accuracy. Inspect the predictions before relying on them.

Try it yourself

1. Choose an image you have permission to use and predict which objects you expect the model to find.
2. Change just one class name or list in groups, run the script again, and inspect the saved image and printed rows.
3. Compare false positives, missed objects and scores. Keep the input and other settings fixed if you want an interpretable comparison.

For a real application, test a representative labelled set covering the relevant objects, viewpoints and conditions. This single synthetic scene is not an accuracy benchmark, a validated real-time system or a surveillance application. The OpenAI CLIP model card discusses research scope, limitations and potential bias.

Troubleshooting

• “Configure this project weights directory first”: run setup_local.py from the same project and then start world_demo.py in a separate process. Re-run setup after moving the project.
• Missing checkpoint or image: inspect the file paths shown above; the script checks them before model initialization.
• CLIP import or installation problem: verify that the tested Ultralytics CLIP fork is installed in this virtual environment, and that the correct kernel or Python executable is used. The separately named PyPI clip project is not a substitute for the specified fork.
• No detection or an unexpected box: check the input, class wording, visible output and threshold. A lower threshold may show additional weak candidates, but does not automatically improve correctness.

Sources and license

Both complete example scripts on this page are © 2026 ARTSCI - Author of the AI/ML/CV course materials, licensed under GNU AGPL-3.0 only, without warranty. Preserve applicable notices and provide corresponding source when sharing covered modifications. This code notice does not automatically license the course image, narration or video.

Ultralytics software and models retain their own terms; see the Ultralytics assets license and licensing information. Ultralytics states that its YOLO-World weights were migrated from the original YOLO-World project, which describes its own code as GPL-3.0. The Ultralytics CLIP fork and original OpenAI CLIP code have separate license texts; consult the upstream sources and model card for the specific components and your use.

To compare this prompt-based approach with adapting a model on labelled images, see how to fine-tune YOLO on your own images. The full English course covers these ideas and other projects: Computer Vision: OpenCV, Machine Learning and YOLO.

How to Train YOLO on Your Own Images: A Step-by-Step Python Tutorial

How can you make an object detector recognize classes from your own labelled images? In this guide we fine-tune a pretrained YOLOv8n model on a small dataset, then inspect a prediction with Python.

Our example uses 12 synthetic training images and 4 synthetic validation images containing boats and buoys. In one local CPU run, the model predicted two boats and one buoy on the selected validation image. This is an educational example: the validation set helped select the checkpoint, and the result does not establish accuracy on independent real-world data.

Watch the tutorial

Watch the full Python and YOLO tutorial, or see the short demonstration.

Actual prediction from the recorded CPU run on val_02.png. This is a validation image, not an independent final test.

What you will build

1. Organize labelled images into training and validation folders.
2. Understand the normalized YOLO bounding-box format.
3. Point a dataset configuration at those folders.
4. Fine-tune a local pretrained model and load the saved checkpoint.
5. Run a prediction and interpret its limitations.

This is fine-tuning: it starts from pretrained weights, rather than initializing a detector from scratch. You can use your own appropriately licensed images and annotations. The particular synthetic fixture shown here is supplied with lesson 29 of my Computer Vision: OpenCV, Machine Learning and YOLO course.

1. Install the dependencies

The recorded experiment ran on Windows 11 using Python 3.13.9, Ultralytics 8.4.116 and PyTorch 2.13.0+cpu. The commands below create a Windows environment and pin the direct Ultralytics dependency; they do not lock every transitive dependency. Other versions, operating systems and hardware may yield different results.

py -3.13 -m venv .venv
.venv\Scripts\python.exe -m pip install --upgrade pip
.venv\Scripts\python.exe -m pip install ultralytics==8.4.116

Installing packages and obtaining the initial checkpoint need an internet connection unless you already have the required files cached. The training demonstration itself used local images and local weights. Your environment's dependencies and model files must be available first.

2. Organize images and annotations

Create this layout next to the training script. Every image has a corresponding label file with the same base name:

custom-yolo/
├── train_yolo.py
├── yolov8n.pt
└── synthetic_harbor_dataset/
    ├── data.yaml
    ├── images/
    │   ├── train/  (e.g. train_00.png)
    │   └── val/    (e.g. val_02.png)
    └── labels/
        ├── train/  (e.g. train_00.txt)
        └── val/    (e.g. val_02.txt)

For example, train_00.png corresponds to train_00.txt. For a different dataset, change the image paths and class names to match your data. Do not place images without permission into your project or public downloads.

Read one YOLO annotation row

Each row describes one object using five values:

class_index x_center y_center width height

The index identifies a class; the remaining four numbers are normalized by the image width or height and typically range from 0 to 1. In this fixture, the first row of labels/train/train_00.txt is:

0 0.271484 0.691406 0.277344 0.187500

Class 0 means boat. The center is approximately (0.271484, 0.691406); its width and height are 0.277344 and 0.187500. The input is 256 × 256 pixels, so the x center is approximately 0.271484 × 256 = 69.5 pixels from the left edge. The rectangle below is computed from this exact row, rather than drawn by eye.

One image and the box defined by the example row. Inspect all objects, not just a single row, before training.

Create and check a label file from pixel coordinates

Here is a small, reproducible way to turn pixel coordinates into a YOLO label. This example uses the synthetic train_00.png image from the course fixture. Its four rectangles contain two boats and two buoys. For this demonstration we reconstructed their pixel corners from the fixture’s existing annotations; they are not newly measured by an annotation tool. For your own image, determine the correct rectangle for every object first—for example using a suitable annotation tool—and use its pixel corners in place of these values.

For each box, left and right are horizontal pixel positions, and top and bottom are vertical positions measured from the upper-left corner. The code divides horizontal values by the image width and vertical values by its height. Class 0 is boat; class 1 is buoy. Use the same numeric mapping in your data.yaml.

Inside the project folder from step 1, create a separate annotation-demo folder and copy the example image to annotation-demo/input/train_00.png. Save the complete script below as annotation-demo/annotate_example.py. This image is 256 × 256 pixels; adapt the dimensions check, filename, class IDs, coordinates and class names to use a different dataset.

custom-yolo/
├── .venv/
└── annotation-demo/
    ├── annotate_example.py
    └── input/
        └── train_00.png

The script uses Pillow, included in the demonstrated environment. If it is not installed in your environment, run .venv\Scripts\python.exe -m pip install pillow from the project folder. Then, still from the project folder, run:

.venv\Scripts\python.exe annotation-demo\annotate_example.py

Complete example source: copyright © 2026 ARTSCI — Author of the AI/ML/CV course materials, licensed under GNU AGPL version 3 only, without warranty. The sample image and other course media are separate assets and are not automatically licensed under this code notice.

"""Create a new YOLO label from pixel corners; never overwrite an existing label."""
from pathlib import Path
from PIL import Image, ImageDraw

image_path = Path(__file__).parent / 'input' / 'train_00.png'
output_dir = Path(__file__).parent / 'output'
image = Image.open(image_path).convert('RGB')
width, height = image.size
assert (width, height) == (256, 256)

# class_id, left, top, right, bottom in pixels. Origin: upper-left.
# These corners were reconstructed from this fixture's existing annotations.
boxes = [
    (0, 34, 153, 105, 201),
    (0, 118, 122, 181, 172),
    (1, 46, 189, 66, 209),
    (1, 186, 149, 212, 175),
]
rows = []
for class_id, left, top, right, bottom in boxes:
    assert class_id in (0, 1) and 0 <= left < right <= width and 0 <= top < bottom <= height
    x_center = (left + right) / (2 * width)
    y_center = (top + bottom) / (2 * height)
    box_width = (right - left) / width
    box_height = (bottom - top) / height
    rows.append(f'{class_id} {x_center:.6f} {y_center:.6f} {box_width:.6f} {box_height:.6f}')

output_dir.mkdir(parents=True, exist_ok=True)
label_path = output_dir / f'{image_path.stem}.txt'
with label_path.open('x', encoding='utf-8', newline='\n') as label_file:
    label_file.write('\n'.join(rows) + '\n')

# Reconstruct the saved boxes for visual verification of the written bytes.
saved_rows = label_path.read_text(encoding='utf-8').splitlines()
assert len(saved_rows) == len(boxes)
overlay = image.copy()
draw = ImageDraw.Draw(overlay)
for row in saved_rows:
    cls, xc, yc, bw, bh = row.split()
    xc, yc, bw, bh = map(float, (xc, yc, bw, bh))
    rect = ((xc-bw/2)*width, (yc-bh/2)*height, (xc+bw/2)*width, (yc+bh/2)*height)
    draw.rectangle(rect, outline='#ffdb40' if cls == '0' else '#5effac', width=2)
overlay.save(output_dir / 'train_00_saved_labels.png')
print('Input:', image_path.name, width, 'x', height)
print('New label:', label_path.name)
print('Saved rows:')
print('\n'.join(saved_rows))
print('Verified overlay:', 'train_00_saved_labels.png')

The script creates annotation-demo/output/train_00.txt with four lines and annotation-demo/output/train_00_saved_labels.png, which draws the rectangles reconstructed from the saved label file. The first line is 0 0.271484 0.691406 0.277344 0.187500. Inspect the overlay and correct any missing or misplaced boxes before using a label for training.

The file is opened in exclusive-create mode to protect existing labels. Running the same demo again with the same output path raises FileExistsError. To repeat it, use a new demo directory or a different output filename; do not overwrite a label without reviewing it. Once your annotations for the entire image are complete and checked, put the label file under the matching labels/train or labels/val directory with the same base name as the image.

3. Create the dataset configuration

For the example layout, save this as synthetic_harbor_dataset/data.yaml:

train: images/train
val: images/val
names:
  0: boat
  1: buoy

The training and validation paths identify separate groups of images. The names mapping defines the meaning of each class index in this custom dataset. Check paths and labels against your actual installed Ultralytics version and project layout; the example was run with the YAML beside these folders.

4. Obtain the starting weights

Download the official YOLOv8n checkpoint as a separate setup step and place it next to train_yolo.py. The file used in the demonstration has SHA-256 f59b3d833e2ff32e194b5bb8e08d211dc7c5bdf144b90d2c8412c47ccfc83b36. Review the model asset license and Ultralytics licensing information for your use. A checksum identifies bytes; it does not by itself grant usage rights.

5. Train and predict with the complete script

Save the following as train_yolo.py in the project folder. It checks for the local checkpoint, dataset YAML and selected validation image, fine-tunes the checkpoint, loads the selected best.pt file, and saves the prediction as prediction_val_02.jpg.

# SPDX-License-Identifier: AGPL-3.0-only
# Copyright (c) 2026 ARTSCI - Author of the AI/ML/CV course materials.
"""Fine-tune a local YOLO checkpoint on an annotated dataset and inspect a validation image."""
from pathlib import Path

import cv2
from ultralytics import YOLO


def main():
    root = Path(__file__).resolve().parent
    dataset = root / "synthetic_harbor_dataset"
    data_yaml = dataset / "data.yaml"
    validation_image = dataset / "images" / "val" / "val_02.png"
    starting_weights = root / "yolov8n.pt"

    for file_path in (data_yaml, validation_image, starting_weights):
        if not file_path.is_file():
            raise FileNotFoundError(f"Required local file: {file_path}")

    model = YOLO(str(starting_weights))
    training = model.train(
        data=str(data_yaml), epochs=50, imgsz=256, batch=4, workers=0,
        device="cpu", project=str(root / "runs"), name="custom_yolo",
        exist_ok=False, verbose=False, plots=False, mosaic=0.0,
        seed=0, deterministic=True,
    )

    best_weights = Path(training.save_dir) / "weights" / "best.pt"
    if not best_weights.is_file():
        raise FileNotFoundError(f"Training did not produce a checkpoint: {best_weights}")
    final_model = YOLO(str(best_weights))
    result = final_model(
        str(validation_image), conf=0.25, iou=0.30, device="cpu", verbose=False,
    )[0]

    print(f"Best checkpoint: {best_weights}")
    for box in result.boxes:
        class_name = result.names[int(box.cls.item())]
        print(f"{class_name}: confidence {float(box.conf.item()):.3f}")
    output = root / "prediction_val_02.jpg"
    if not cv2.imwrite(str(output), result.plot(line_width=1, font_size=8)):
        raise OSError(f"Could not save prediction image: {output}")
    print(f"Saved validation-image prediction: {output}")


if __name__ == "__main__":
    main()

Complete sample source above: copyright © 2026 ARTSCI — Author of the AI/ML/CV course materials, licensed under GNU AGPL version 3, without warranty. Preserve applicable notices and provide corresponding source when sharing covered modifications. Ultralytics software and model weights retain their own terms. Dataset images, course media and other independent assets are not automatically licensed under this code notice.

From the project folder, run:

.venv\Scripts\python.exe train_yolo.py

data selects the YAML. epochs=50 means 50 passes through the training set; imgsz=256 selects the processing size, batch=4 controls batch size, and device="cpu" chooses the CPU. The seed and deterministic setting help reproduce this particular example, but they do not guarantee bit-identical results across every environment. Training duration depends on the computer.

6. Inspect the actual result

In the recorded run, prediction on val_02.png at conf=0.25 and iou=0.30 returned two boat boxes and one buoy box. These matched the three labelled classes in that particular image. The checkpoint and image shown in the video belong to that run; a new run can produce different confidences or boxes.

A box confidence is the model's score for that prediction, not a measurement of how often the system will be correct in real life. The four validation images were used during training and checkpoint selection. Their prediction and validation metrics are therefore not an independent final test.

7. Troubleshoot the common problems

• Required local file is missing: check the project layout and filenames, including the checkpoint and data.yaml. For your own dataset, change validation_image to an existing image.
• Images or labels are not found: check matching filename stems and whether train and val resolve to the intended folders.
• Boxes look wrong: check that the annotations use normalized center coordinates and width/height, rather than pixel corners. Draw them on the source image before training.
• No box appears: inspect your labels, classes and input, then investigate the confidence threshold and training output. Lowering a threshold changes which predictions appear; it does not guarantee correct detections.
• Training is slow: CPU runtime depends on your hardware, configuration and dataset. Reduce or adjust a test run deliberately, and report the actual settings when comparing results.

What to do next

Try adding an image you have permission to use and manually checking every annotation. For a real application, collect representative examples across cameras and lighting, prevent similar images from leaking between splits, inspect errors, and evaluate on a separate test set that was not used for training or checkpoint selection. A tiny synthetic fixture teaches the steps; it does not establish production reliability.

For another application of these ideas, read how a Parking Space Counter assigns detections to spaces. You can explore more practical projects in my Computer Vision: OpenCV, Machine Learning and YOLO course.

Parking Space Counter with Python and YOLO: From Detections to Occupancy

Finding vehicles is only half of a parking-space counter. The application must also decide which detected vehicle belongs to which monitored space.

In this Python and YOLO example, six calibrated regions produce four occupied and two free statuses. Raising one application threshold changes the result to three occupied and three free, even though the image has not changed. Following that change back to one detection makes the rule much easier to understand.

We use a synthetic 1024 × 1024 course scene and a local Windows/CPU run. This is a static-image demonstration, not a live-camera reliability or accuracy benchmark. Other partial vehicles appear outside the six monitored regions; the reported total is about those six spaces, not every vehicle in the image.

Watch the Parking Lot Monitor walkthrough

Watch the complete Python and YOLO tutorial, or see the quick parking-space demonstration.

Red regions are occupied and green regions free according to the rule. Yellow rectangles and points show accepted detector boxes and their centers. The regions approximate perspective bays rather than tracing their painted outlines.

The instructional figures include extra yellow detection boxes and centers to make the decisions inspectable. The minimal script below saves the colored region-and-ID overlays and prints the assignment rows; it does not add those extra yellow annotations to its saved images.

The pipeline: image, detections, regions, decisions

1. Load one explicit local image, checkpoint and region calibration.
2. Run YOLOv8n once with a candidate confidence floor of 0.10.
3. Keep vehicle classes: car, truck and bus.
4. Filter those fixed candidates using the parking acceptance threshold.
5. Assign accepted box centers to the calibrated rectangles.
6. Count the resulting occupied and free rows.

YOLO supplies predictions; our Python rule supplies the occupancy decision. A free row means that no accepted candidate center matched that region. It does not independently verify that the physical space is empty.

Prepare the local inputs

Save the complete script below as parking.py. Place yolov8n.pt, parking_lot_sample.png and parking_zones.json beside it. The demonstrated input and its calibration are included in lesson 50 of my course. You can instead use your own permitted image, but you must calibrate regions for that image.

The tested checkpoint matches the official YOLOv8n release asset, SHA-256 f59b3d833e2ff32e194b5bb8e08d211dc7c5bdf144b90d2c8412c47ccfc83b36. Obtain that file as a separate setup step and review the model license and Ultralytics licensing options. The script checks that the required local files exist before loading the model; it does not download missing inputs.

For Windows with Python 3.13 installed:

py -3.13 -m venv .venv
.venv\Scripts\python.exe -m pip install ultralytics==8.4.24 numpy==2.3.5 opencv-python==4.13.0.92
.venv\Scripts\python.exe parking.py

The measured environment used Python 3.13.9, Ultralytics 8.4.24, PyTorch 2.10.0, NumPy 2.3.5 and OpenCV 4.13.0.92. The installation command pins three direct dependencies, not the complete transitive environment. This is not a fresh-install claim for every operating system. Installation and obtaining the model may need internet access; the demonstrated processing runs offline with local files.

Calibrate regions in the input image's coordinates

Here is the demonstrated calibration. Save it as parking_zones.json. Coordinates are [left, top, right, bottom] in the original 1024 × 1024 image:

{
  "image_size": [1024, 1024],
  "zones": [
    {"spot_id": "A1", "rect": [146, 55, 399, 337]},
    {"spot_id": "A2", "rect": [402, 55, 649, 337]},
    {"spot_id": "A3", "rect": [651, 55, 856, 337]},
    {"spot_id": "B1", "rect": [72, 570, 373, 978]},
    {"spot_id": "B2", "rect": [375, 570, 696, 978]},
    {"spot_id": "B3", "rect": [698, 570, 959, 978]}
  ]
}

These coordinates belong to this image and viewpoint. Equal dimensions alone do not make them valid for another camera.

load_zones rejects a calibration with different image dimensions, an empty region list, duplicate identifiers, invalid rectangles, or overlapping regions. Those checks catch structural mistakes. They do not prove that a rectangle corresponds to the correct physical bay. Inspect the overlay, especially after changing the image or viewpoint.

Complete runnable parking-space counter

# SPDX-License-Identifier: AGPL-3.0-only
# Copyright (c) 2026 ARTSCI - Author of the AI/ML/CV course materials.
"""Assign local YOLO candidates to calibrated parking spaces; no asset downloads."""
from pathlib import Path
import json
import sys
import cv2
import numpy as np
from ultralytics import YOLO


def load_zones(path, shape):
    payload = json.loads(Path(path).read_text(encoding="utf-8"))
    if payload["image_size"] != [shape[1], shape[0]]:
        raise ValueError("Zone calibration dimensions differ from the selected image.")
    zones = payload["zones"]
    if not zones or len({z["spot_id"] for z in zones}) != len(zones):
        raise ValueError("Use nonempty zones with unique spot IDs.")
    for i, zone in enumerate(zones):
        x1, y1, x2, y2 = zone["rect"]
        if not 0 <= x1 < x2 <= shape[1] or not 0 <= y1 < y2 <= shape[0]:
            raise ValueError("Invalid or out-of-image zone.")
        for other in zones[:i]:
            a, b, c, d = other["rect"]
            if max(x1, a) < min(x2, c) and max(y1, b) < min(y2, d):
                raise ValueError("Overlapping zones could assign one center twice.")
    return zones


def detect_candidates(model, image):
    result = model.predict(
        source=image, conf=0.10, imgsz=640, device="cpu", iou=0.70,
        max_det=300, rect=True, half=False, classes=None,
        augment=False, agnostic_nms=False, verbose=False,
    )[0]
    detections = []
    for box in result.boxes:
        label = result.names[int(box.cls.item())]
        if label not in {"car", "truck", "bus"}:
            continue
        x1, y1, x2, y2 = box.xyxy[0].cpu().tolist()
        detections.append(dict(
            label=label, confidence=float(box.conf.item()), box=[x1, y1, x2, y2],
            center=[(x1 + x2) / 2, (y1 + y2) / 2],
        ))
    return detections


def assign_spots(detections, zones, confidence=0.20):
    accepted = [d for d in detections if d["confidence"] >= confidence]
    rows = []
    for zone in zones:
        x1, y1, x2, y2 = zone["rect"]
        matches = [d for d in accepted
                   if x1 <= d["center"][0] < x2 and y1 <= d["center"][1] < y2]
        best = max(matches, key=lambda d: d["confidence"]) if matches else None
        rows.append(dict(spot_id=zone["spot_id"], status="occupied" if best else "free",
                         confidence=round(best["confidence"], 3) if best else None,
                         center=best["center"] if best else None))
    return rows


def draw_status(image, zones, rows):
    canvas = image.copy()
    states = {row["spot_id"]: row["status"] for row in rows}
    for zone in zones:
        x1, y1, x2, y2 = zone["rect"]
        color = (70, 170, 70) if states[zone["spot_id"]] == "free" else (50, 70, 225)
        cv2.rectangle(canvas, (x1, y1), (x2, y2), color, 4)
        cv2.putText(canvas, zone["spot_id"], (x1 + 8, y1 + 32),
                    cv2.FONT_HERSHEY_SIMPLEX, 1, color, 3)
    return canvas


def main():
    root = Path(__file__).resolve().parent
    image_path = Path(sys.argv[1]) if len(sys.argv) > 1 else root / "parking_lot_sample.png"
    zones_path = Path(sys.argv[2]) if len(sys.argv) > 2 else root / "parking_zones.json"
    weights = root / "yolov8n.pt"
    for path in (image_path, zones_path, weights):
        if not path.is_file():
            raise FileNotFoundError(f"Required local file: {path}")
    image = cv2.imdecode(np.fromfile(image_path, dtype=np.uint8), cv2.IMREAD_COLOR)
    if image is None:
        raise ValueError(f"Cannot decode image: {image_path}")
    zones = load_zones(zones_path, image.shape)
    detections = detect_candidates(YOLO(str(weights)), image)
    output = root / "output"
    output.mkdir(exist_ok=True)
    for confidence in (0.20, 0.30):
        rows = assign_spots(detections, zones, confidence)
        occupied = sum(row["status"] == "occupied" for row in rows)
        print(f"threshold={confidence:.2f}: {occupied} occupied, {len(rows)-occupied} free")
        for row in rows:
            print(f"  {row['spot_id']}: {row['status']} (score={row['confidence']})")
        if not cv2.imwrite(str(output / f"parking_{confidence:.2f}.png"),
                           draw_status(image, zones, rows)):
            raise OSError("Could not save parking overlay.")


if __name__ == "__main__":
    main()

Complete sample source above: copyright © 2026 ARTSCI — Author of the AI/ML/CV course materials, licensed under GNU AGPL version 3, without warranty. Keep applicable notices and provide corresponding source when sharing covered modifications. Ultralytics software and weights retain their own terms. Independent course images, narration, slides and videos are not automatically licensed under AGPL.

The script's default paths are relative to the saved Python file. To select your own image and its matching calibration, pass both paths:

.venv\Scripts\python.exe parking.py "C:\images\my_parking.png" "C:\images\my_zones.json"

Keep detector inference separate from the parking threshold

detect_candidates runs prediction once, with conf=0.10, imgsz=640, CPU and the other shown settings fixed. In this execution it returns eight vehicle candidates. Some are outside the monitored spaces, and two predictions can describe the same region. A candidate count is not a parking-space count.

The loop in main compares application thresholds 0.20 and 0.30 by filtering that same post-NMS candidate set. It does not rerun YOLO or non-maximum suppression. This controlled comparison keeps the image, model and underlying predictions unchanged. The official prediction documentation describes confidence and overlap-suppression settings; our additional acceptance threshold belongs to the application rule.

Assign the center, then count the space

For each box, the center is ((x1+x2)/2, (y1+y2)/2). assign_spots tests that center against each region, including left/top boundaries and excluding right/bottom boundaries. With non-overlapping regions, a point cannot belong to two spaces. If several accepted centers match one space, the rule keeps the highest-scoring match and still counts only one occupied row.

This is a center-in-rectangle rule, not an intersection-over-union or segmentation calculation. A long, partly occluded or incorrectly localized vehicle may not behave as expected. Test the geometry and the detector output separately.

Inspect the actual output

threshold=0.20: 4 occupied, 2 free
  A1: occupied (score=0.771)
  A2: occupied (score=0.428)
  A3: free (score=None)
  B1: occupied (score=0.222)
  B2: occupied (score=0.501)
  B3: free (score=None)
threshold=0.30: 3 occupied, 3 free
  A1: occupied (score=0.771)
  A2: occupied (score=0.428)
  A3: free (score=None)
  B1: free (score=None)
  B2: occupied (score=0.501)
  B3: free (score=None)

At 0.20, A1, A2, B1 and B2 are occupied; A3 and B3 are free according to the rule. Seven candidate detections pass that threshold, but only four monitored spaces receive a match. The detector labels the red vehicle in B2 as bus with score about 0.501. The application accepts that class as a vehicle, so a useful occupancy result here does not establish perfect vehicle classification.

Why did B1 become free when the car stayed?

The B1 candidate has score approximately 0.22236 and center (229.74, 740.91). It passes 0.20, but fails 0.30. B1 therefore loses its accepted center and changes to free. Five candidates pass the higher threshold, while only three monitored spaces receive matches.

Inspecting one candidate explains the changed decision. The center coordinates refer to the original full image, not this crop's coordinate system.

Same pixels, same calibration and same detector candidate set. Only the application acceptance threshold changed.

A stricter threshold can remove mistakes, but it can also miss occupied spaces. A confidence score is not calibrated certainty, and a higher threshold does not automatically make a counter more accurate.

What to verify before adapting the example

Use permitted inputs, recalibrate the regions, and inspect each selected vehicle center alongside the assignment table. Test weak detections, occlusion, shadows and boundaries. Check representative labelled examples before relying on physical occupancy claims. This single synthetic scene does not measure field accuracy or live-camera performance.

For related projects, see how to read YOLO detection outputs and why track IDs are not vehicle counts. My English course, Computer Vision: OpenCV, Machine Learning and YOLO, develops the full Parking Lot Monitor application alongside other practical projects.

Python Document Scanner: Detect Corners and Correct Perspective

A tilted page can become a rectangular scan without training a neural network. The useful idea is to find a plausible page boundary, keep its four corners in a consistent order, and map them onto a rectangle.

This walkthrough builds a small Python document scanner with OpenCV. It runs locally on a CPU and saves both a color result and an optional black-and-white view. There is no model checkpoint and no OCR. The example uses an explicitly synthetic document scene from my course, so a clean result here is a demonstration of the pipeline rather than evidence that every photograph will work.

Watch the working example

Watch the complete Python and OpenCV scanner walkthrough, or see the quick before-and-after demonstration.

Same input pixels, transformed using four detected corners. The result is 740 × 523 pixels in this execution.

What does the scanner actually do?

1. Convert the image to grayscale and blur small variations.
2. Use Canny edges and morphological closing to prepare boundaries.
3. Select the largest qualifying convex four-corner contour.
4. Order those corners and calculate a perspective transform.
5. Save the flattened color image and an optional thresholded view.

OpenCV's contour documentation explains how contours represent boundaries in a binary image. A boundary is a shape, not an understanding of the object. That distinction is the main limitation to keep in mind.

Prepare Python, OpenCV and one local image

Save the full code below as scanner.py. Place document_scene.png beside it to use the demonstrated course input, available in lesson 53. Alternatively, supply the path to an image you have permission to use. The example does not download or silently replace your image.

On Windows with Python 3.13 installed, create a separate environment and install the two demonstrated package versions:

py -3.13 -m venv .venv
.venv\Scripts\python.exe -m pip install numpy==2.3.5 opencv-python==4.13.0.92
.venv\Scripts\python.exe scanner.py

To process your own file instead of the default input:

.venv\Scripts\python.exe scanner.py "C:\images\my_page.jpg"

The measured execution used Python 3.13.9, NumPy 2.3.5 and OpenCV 4.13.0.92 on Windows. These commands describe that setup; they are not a fresh-install test of every operating system. Package installation may require internet access. Once the packages and image are present, this scanner needs no network connection, model download or GPU.

Complete runnable scanner

# Copyright (c) 2026 ARTSCI - Author of the AI/ML/CV course materials.
"""Straighten a page candidate with local OpenCV operations; no trained model."""
from pathlib import Path
import sys
import cv2
import numpy as np


def edge_map(image, low=50, high=150):
    if not 0 <= low < high:
        raise ValueError("Require 0 <= low < high.")
    gray = cv2.cvtColor(image, cv2.COLOR_BGR2GRAY)
    blurred = cv2.GaussianBlur(gray, (5, 5), 0)
    edges = cv2.Canny(blurred, low, high)
    return cv2.morphologyEx(edges, cv2.MORPH_CLOSE, np.ones((5, 5), np.uint8))


def find_page(edges):
    contours, _ = cv2.findContours(edges, cv2.RETR_LIST, cv2.CHAIN_APPROX_SIMPLE)
    for contour in sorted(contours, key=cv2.contourArea, reverse=True):
        perimeter = cv2.arcLength(contour, True)
        candidate = cv2.approxPolyDP(contour, 0.02 * perimeter, True)
        large = cv2.contourArea(candidate) >= 0.10 * edges.size
        if len(candidate) == 4 and large and cv2.isContourConvex(candidate):
            return candidate.reshape(4, 2).astype(np.float32)
    return None


def order_corners(points):
    points = np.asarray(points, dtype=np.float32).reshape(4, 2)
    if not np.isfinite(points).all() or len(np.unique(points, axis=0)) != 4:
        raise ValueError("Four distinct finite corners are required.")
    center = points.mean(axis=0)
    angles = np.arctan2(points[:, 1] - center[1], points[:, 0] - center[0])
    ordered = points[np.argsort(angles)]
    if not cv2.isContourConvex(ordered) or abs(cv2.contourArea(ordered)) < 1:
        raise ValueError("Corners must form a non-degenerate convex quadrilateral.")
    start = np.argmin(ordered.sum(axis=1))
    return np.roll(ordered, -start, axis=0)


def flatten_page(image, corners):
    tl, tr, br, bl = corners
    width = max(2, round(max(np.linalg.norm(tr - tl), np.linalg.norm(br - bl))))
    height = max(2, round(max(np.linalg.norm(bl - tl), np.linalg.norm(br - tr))))
    target = np.float32([[0, 0], [width-1, 0], [width-1, height-1], [0, height-1]])
    matrix = cv2.getPerspectiveTransform(np.float32(corners), target)
    return cv2.warpPerspective(image, matrix, (width, height))


def document_mode(photo):
    gray = cv2.cvtColor(photo, cv2.COLOR_BGR2GRAY)
    blurred = cv2.GaussianBlur(gray, (3, 3), 0)
    normalized = cv2.normalize(blurred, None, 0, 255, cv2.NORM_MINMAX)
    return cv2.adaptiveThreshold(normalized, 255, cv2.ADAPTIVE_THRESH_GAUSSIAN_C,
                                 cv2.THRESH_BINARY, 31, 15)


def main():
    root = Path(__file__).resolve().parent
    source = Path(sys.argv[1]) if len(sys.argv) > 1 else root / "document_scene.png"
    if not source.is_file():
        raise FileNotFoundError(f"Required local image: {source}")
    image = cv2.imdecode(np.fromfile(source, dtype=np.uint8), cv2.IMREAD_COLOR)
    if image is None:
        raise ValueError(f"Cannot decode image: {source}")
    edges = edge_map(image)
    candidate = find_page(edges)
    if candidate is None:
        raise ValueError("No qualifying page candidate. Check the image and thresholds.")
    corners = order_corners(candidate)
    photo = flatten_page(image, corners)
    binary = document_mode(photo)
    output = root / "output"
    output.mkdir(exist_ok=True)
    for name, pixels in [("edges", edges), ("photo", photo), ("document", binary)]:
        if not cv2.imwrite(str(output / f"{name}.png"), pixels):
            raise OSError(f"Could not save {name}.png")
    print("Corners [TL, TR, BR, BL]:", corners.astype(int).tolist())
    print(f"Scan: {photo.shape[1]} x {photo.shape[0]} pixels")
    print("Saved edges.png, photo.png and document.png in output/")


if __name__ == "__main__":
    main()

Sample code © 2026 ARTSCI — Author of the AI/ML/CV course materials. OpenCV and NumPy retain their own licenses. OpenCV 4.5 and later use Apache License 2.0; NumPy's license is documented separately. This example uses no Ultralytics software or model weights.

Why grayscale, blur and Canny come first

edge_map converts the input to grayscale and applies a 5 × 5 Gaussian blur. Canny uses thresholds 50 and 150 in this run. Closing with a 5 × 5 kernel can connect small gaps, making a boundary easier to follow. These values are starting parameters, not settings that are guaranteed to work for all lighting and backgrounds.

The edge image includes text and background details as well as the page boundary. find_page sorts contours by area, approximates each polygon with a tolerance of 2% of its perimeter, then accepts the first convex candidate with four corners and at least 10% of the image area. Always inspect that selected boundary before interpreting the output as a document scan.

Detected corners for this input: TL=(199,156), TR=(818,127), BR=(884,643), BL=(145,676). The yellow outline is the selected candidate.

Corner order controls the perspective mapping

The source and destination corners must correspond. Here the order is top left, top right, bottom right, bottom left. The code sorts points around their center and rotates that sequence so the point with the smallest x+y comes first. It also rejects duplicate, nonfinite and degenerate corner sets.

This ordering assumes a roughly upright page. It does not identify which way the text should face. Strongly rotated or ambiguous pages need additional orientation handling. Mixing corner correspondence can twist or flip the output even when all four points lie near a page.

cv2.getPerspectiveTransform calculates the mapping from the four input points to the four rectangle points. cv2.warpPerspective resamples the image using that mapping. See the official perspective transformation tutorial for the source-to-destination point relationship.

Read the actual output

Corners [TL, TR, BR, BL]: [[199, 156], [818, 127], [884, 643], [145, 676]]
Scan: 740 x 523 pixels
Saved edges.png, photo.png and document.png in output/

The printed 740 × 523 dimensions are a result of this image and the edge-length rule in flatten_page. They are not a known physical paper size. If the real page dimensions matter, supply that prior information and choose the target aspect ratio accordingly.

Flattened color output. Perspective correction does not recover blurred characters or remove a curved page's three-dimensional shape.

Black and white is an optional view, not OCR

document_mode converts the flattened image to grayscale, blurs it, normalizes its intensity range and applies adaptive Gaussian thresholding with a 31-pixel neighborhood and C=15. A local threshold can help with brightness variation, but it can also discard fine detail or amplify noise. Compare this output with the color scan before deciding which one to keep.

A document-style image. No text has been recognized or exported as characters.

When a convincing scan can still be wrong

A large picture frame, desk mat or other convex four-corner object can satisfy exactly the same rule as a page. In a separate controlled test, the algorithm accepted a plain large rectangle that was not a document. A blank image produced no candidate. Those outcomes are useful checks, but neither establishes general document-detection accuracy.

For your own input, inspect the selected outline, check all four corners and compare the output with the original. Test shadows, low-contrast borders, clutter and rotation. If the script reports no qualifying page, investigate the input and parameters; do not silently crop an arbitrary region and call it a successful scan.

This project demonstrates classical image geometry. If you want to compare it with learned detection, see my first working YOLO object-detection example. My English course, Computer Vision: OpenCV, Machine Learning and YOLO, develops these operations through notebooks and practical applications, including the complete Document Scanner project.

Object Detection in Python with YOLO: Boxes, Classes and Confidence

An object detector gives you more than a name. It estimates where objects are, assigns class labels and reports scores. The useful first project is to connect those predictions to a visible image, then change one setting and inspect the result.

In this example, YOLOv8n processes a synthetic 1024 × 1024 harbor image on a Windows CPU. With the same image and model, it returns 9 boxes at confidence 0.15 and 4 boxes at 0.70. Those are retained predictions, not independently verified counts of physical boats.

Watch the Python walkthrough

Watch the full tutorial on YouTube · Watch the confidence-threshold Short

What does object detection return?

Image classification answers a question about an image's category. Object detection adds estimated locations. In this run, every returned class label is boat, but each prediction has its own score and rectangle. The official detection documentation describes these outputs.

A bounding box does not trace an object's exact outline. It gives four edges around an estimated region. A confidence score helps filter predictions; it is not a certificate that a box is correct. Neither a convincing overlay nor a large number of boxes establishes model accuracy.

Actual low-threshold result. Numbered labels identify returned prediction rows, not persistent object IDs. Overlapping boxes and a questionable region near the seawall are reasons to inspect the output.

Prepare two local files and a small script

Use a local yolov8n.pt checkpoint and an image named harbor_scene.jpg, placed beside detect.py. The demonstrated input and checkpoint are included in lesson 28 of the course. For your own experiment, use an image you have permission to process and update the image path in the script. Your results will depend on that image; the numbers below belong to the demonstrated input.

The checkpoint in this run matches the official Ultralytics YOLOv8n release asset, SHA-256 f59b3d833e2ff32e194b5bb8e08d211dc7c5bdf144b90d2c8412c47ccfc83b36. Downloading the model is a separate setup step; the example deliberately stops if a local file is missing. Review the model license and Ultralytics licensing information before integrating it into another project.

For Windows with Python 3.13 installed, create an isolated environment and install the demonstrated Ultralytics version. Save the following script and put the two assets beside it before running the final command:

py -3.13 -m venv .venv
.venv\Scripts\python.exe -m pip install ultralytics==8.4.24
.venv\Scripts\python.exe detect.py

The measured run used Python 3.13.9, Ultralytics 8.4.24, PyTorch 2.10.0, NumPy 2.3.5 and OpenCV 4.13.0.92. The short installation command pins Ultralytics, not every transitive dependency. It is a starting setup, not a claim that a fresh installation on every machine reproduces the complete tested environment. Installing packages or obtaining assets may need internet access; prediction can run offline once they are present. No GPU or model training is required for this example.

Run the complete Python example

# SPDX-License-Identifier: AGPL-3.0-only
# Copyright (c) 2026 ARTSCI - Author of the AI/ML/CV course materials.
"""Run detection on two local files; do not download missing assets."""
from pathlib import Path
from ultralytics import YOLO

ROOT = Path(__file__).resolve().parent
weights = ROOT / "yolov8n.pt"
image = ROOT / "harbor_scene.jpg"
for path in (weights, image):
    if not path.is_file():
        raise FileNotFoundError(f"Required local file: {path}")

model = YOLO(str(weights))
for confidence in (0.15, 0.70):
    result = model.predict(
        source=str(image), conf=confidence, iou=0.70,
        imgsz=640, device="cpu", max_det=300,
        classes=None, rect=True, half=False,
        augment=False, agnostic_nms=False, verbose=False,
    )[0]
    print(f"conf={confidence:.2f}: {len(result.boxes)} detections")
    for box in result.boxes:
        label = result.names[int(box.cls.item())]
        score = float(box.conf.item())
        xyxy = [round(float(v), 1) for v in box.xyxy[0].tolist()]
        print(f"  {label}: score={score:.3f}, xyxy={xyxy}")
    result.save(filename=str(ROOT / f"detections_{confidence:.2f}.jpg"))

Complete sample source above: copyright © 2026 ARTSCI — Author of the AI/ML/CV course materials. Licensed under GNU AGPL version 3, without warranty. Preserve applicable notices when sharing modified code. Ultralytics software and weights retain their own terms. This code notice does not automatically license independent images, narration or course videos.

Path(__file__).resolve().parent makes the asset paths relative to the saved script. Use it in a Python file; an interactive notebook needs paths appropriate to its working directory instead. The existence checks prevent an accidental missing-model download. YOLO(str(weights)) loads the local checkpoint, and model.predict(...)[0] selects the result for this single image.

The loop runs twice, changing only conf. It keeps IoU at 0.70, processing size at 640, the device on CPU and the other shown options fixed. For this YOLOv8 model, IoU controls overlap suppression. It is a different setting from confidence, so changing both at once would make this comparison harder to interpret. See the prediction argument documentation for their roles.

Read one prediction all the way back to the image

result.boxes contains the retained boxes. The class identifier indexes result.names to obtain a readable label. box.conf supplies the score. box.xyxy supplies left, top, right and bottom coordinates in the original input image's pixel coordinate system.

One prediction over the far-left red boat has label boat, score approximately 0.621 and coordinates [60.3, 517.6, 339.8, 852.8]. The rounded rectangle begins near x = 60 and y = 518, ending near x = 340 and y = 853. Resizing an illustration for display does not turn these into display-pixel coordinates; a custom overlay must apply the image's display transform once.

A crop used to inspect one actual prediction. Its coordinates still refer to the original 1024 × 1024 input.

Here is an abbreviated excerpt from the printed output. The ellipses omit other rows; the script prints every returned row:

conf=0.15: 9 detections
  ...
  boat: score=0.621, xyxy=[60.3, 517.6, 339.8, 852.8]
  ...
conf=0.70: 4 detections

Why does a visible boat lose its box?

The selected score, about 0.621, passes 0.15 but falls below 0.70. In the second run that prediction is absent. The underlying image has not changed, and the visible boat has not disappeared. The acceptance threshold has changed.

Same input, checkpoint, IoU and processing size; higher confidence threshold. Four boxes remain in this execution.

Keeping fewer predictions can remove weak duplicates or mistakes, but it can also remove useful detections. Keeping more can reveal weak candidates while introducing extra errors. We have not annotated every true object in this synthetic scene, so this is a controlled output comparison, not an accuracy benchmark. A score of 0.621 should not be read as a demonstrated 62.1% probability of correctness.

The two saved images make the result inspectable. For another experiment, retain the original input, record the model and versions, and change one parameter at a time. Test representative images and decide which missed or false detections matter for your application.

For the next step, see how detection becomes tracking and crossing events in Traffic Counter. The full notebooks, supporting files and complete applications are developed in my English course, Computer Vision: OpenCV, Machine Learning and YOLO.

A Knee Angle Is Not a Repetition Counter

A knee angle is a measurement. A repetition count is a decision about a sequence. Connect them without defining that decision and a plausible skeleton can give a misleading impression of what the camera observed.

This synthetic demonstration contains 241 frames at 24 frames per second. Our pose pipeline records two bottom-to-standing returns, at about 1.79 and 8.17 seconds. Yet the first frame already shows the person down. Only the second event has a complete standing–down–standing cycle visible within the recording.

Watch the pose-counter breakdown

Watch the full breakdown on YouTube · Watch the short demonstration

The first frame, with predicted landmarks. The recording starts at the bottom, so it cannot show the descent that preceded the first return.

Follow the input all the way to an event

The input is a 1280 × 720 video, analysed offline with YOLOv8n-pose on a CPU. Pose estimation supplies predicted landmark coordinates and scores. The official Ultralytics pose documentation describes these keypoint outputs. The model does not decide what our application should call a repetition.

This example accepts exactly one detected pose per frame. For each leg, it examines the hip, knee and ankle scores, then chooses the leg with the higher minimum score. That minimum must reach 0.15, and all three coordinates must be finite. This filter selects observations; passing it does not establish landmark correctness.

The hip–knee and ankle–knee vectors meet at the knee. Their normalized dot product and arccosine produce an estimated image-space angle. It depends on the camera view and predicted points. An angle alone contains no memory of whether the person was previously standing or down.

Before counting, the implementation smooths the angles with a centered nine-frame moving average. It pads segment edges and processes each continuous valid segment separately, leaving missing observations as gaps. Because a centered filter uses future frames, this is an offline example. A live version needs a deliberate causal filter and new checks for delay and changed event timing.

Remember a bottom, then accept a return

The counter starts unarmed. A finite angle at or below 70° arms it. A later armed observation at or above 150° records the current frame and disarms it. Another standing frame adds nothing. A missing angle clears the remembered state.

The separated thresholds provide hysteresis: movement around the standing threshold cannot repeatedly count without another bottom observation. These are demonstration settings for this signal, not a definition of correct exercise technique. Here is the complete counting rule; it expects angles that have already been measured and smoothed. Represent a missing numeric angle as float("nan"), rather than passing None or a pose record:

import math


def count_returns(angles, down=70.0, up=150.0):
    if not 0 < down < up <= 180:
        raise ValueError("Require 0 < down < up <= 180.")
    armed, events = False, []
    for frame, angle in enumerate(angles):
        if not math.isfinite(angle):
            armed = False
        elif angle <= down:
            armed = True
        elif armed and angle >= up:
            events.append(frame)
            armed = False
    return events

Code adapted from the course practical source. Copyright © 2026 ARTSCI — Author of the AI/ML/CV course materials. Licensed under GNU AGPL version 3, without warranty. This code notice does not automatically license independent images, audio, video or course presentation. Ultralytics software and model terms remain applicable.

Inspect the frame that changes the count

All 241 frames in this run provide usable leg observations. At frame 42, the smoothed angle is about 148.5°. At frame 43, it is about 150.9°. The counter was armed by an earlier bottom observation, so frame 43 satisfies the upper threshold and records the first return.

Frame 43: the estimated leg points come from the recorded pose output; the displayed 150.9° is the smoothed angle used by the rule.

Accepted return	Frame index	Time
First	43	1.79 s
Second	196	8.17 s

Frame indices start at zero; the times are frame index divided by 24. These are two detections under this rule for this particular clip. They are not an accuracy score against independently labelled movement.

Test the meaning, not just the final total

Run these small examples after the function. They separate state-machine behavior from pose-model quality and make the recording-boundary issue explicit:

# Standing -> bottom -> standing: one accepted return.
assert count_returns([160, 60, 100, 150, 170]) == [3]
# Starting down still counts; it does not prove a full observed cycle.
assert count_returns([60, 150, 160]) == [1]
# A shallow bend never arms the rule.
assert count_returns([160, 145, 160]) == []
# Missing data clears the remembered bottom.
assert count_returns([60, float("nan"), 150]) == []
# Both thresholds are inclusive; remaining high does not count again.
assert count_returns([70, 149.9, 150, 170]) == [2]

The second test matters most here. A clip that begins down immediately arms this rule. Its next return counts even though no preceding descent was recorded. If your requirement is a fully observed cycle, begin with an observed standing state, then require a bottom and a return. That is a different state machine; it should reject the initial partial cycle.

Thresholds also affect timing. Keeping this clip and smoothed signal fixed but raising the upper threshold to 160° moves the events to frames 48 and 200. That change does not by itself establish better accuracy. Inspect which observations satisfy the rule before interpreting a changed total or timestamp.

The same distinction appears in the traffic-counter example: track IDs and accepted crossings answer different questions. For pose data, define the event, inspect recording boundaries and missing observations, then explain each accepted transition. The full notebook, local application and broader practical workflow are developed in my English course, Computer Vision: OpenCV, Machine Learning and YOLO.

Count Cars with YOLO: Why Track IDs Are Not Vehicle Counts

A box follows a car, a number appears above it, and a counter increases. It looks simple. But what exactly did the application count: detections, tracking identities, or vehicles passing a particular point? Those are different measurements.

In this Python example, YOLOv8n detects vehicles, ByteTrack associates observations across frames, and a separate rule accepts downward crossings of a horizontal line. Keeping those responsibilities separate makes the result easier to explain and test.

Watch the tutorial

Watch the full tutorial on YouTube · Watch the short demonstration

The demonstration uses a synthetic traffic scene. Boxes, IDs and crossing events come from the same recorded tracking run.

Three measurements that should stay separate

A detection describes an object in one frame. A track ID associates observations over time. A crossing event records an observation sequence that satisfies our counting rule.

One car can appear in many frames. A tracker may also lose a vehicle and assign a new ID when it is detected again. Counting every detection would repeatedly count visible objects; counting distinct IDs would measure the tracker's associations. Neither automatically tells us how many physical vehicles passed our chosen location.

In this particular 241-frame, 1280 × 720 clip, the recorded run contains 23 temporary IDs. Our selected rule accepts 7 crossing events. This is a result for one clip and configuration, not an accuracy claim for general traffic footage.

Define a crossing before calculating a total

Image coordinates start at the top left: y increases downward. We place the line at line_y = 413 and restrict its horizontal span to x_range = (80, 1000). A qualifying center must move from above the line to the line or below it. Its current x must fall within that span.

The complete rule below remembers the previous y for each ID. It also remembers IDs already counted during this function call.

# Turn tracked positions into one downward pass per ID.
def count_passes(frame_tracks, line_y, x_range):
    previous_y, counted_ids, events = {}, set(), []
    for frame in frame_tracks:
        for obj in frame["objects"]:
            track_id = obj["track_id"]
            x, y = obj["center"]
            crossed = previous_y.get(track_id, y) < line_y <= y
            if crossed and x_range[0] <= x <= x_range[1] and track_id not in counted_ids:
                counted_ids.add(track_id)
                events.append({"frame": frame["frame_idx"], "id": track_id, "class": obj["class_name"]})
            previous_y[track_id] = y
    return events

Code: Copyright © 2026 ARTSCI — Author of the AI/ML/CV course materials. Licensed under the GNU AGPL version 3, without warranty. This code notice does not automatically apply to the independent images, video or course presentation.

The decisive comparison is previous_y.get(track_id, y) < line_y <= y. For a first observation, the default previous value is the current y, so the comparison is false. An object first detected below the line therefore does not produce an invented crossing.

counted_ids prevents a second accepted event for the same ID in one call. The function updates the previous y after examining every observation, even when the object is outside the horizontal span. It does not interpolate where the center crossed the line: the span check uses the current x.

Explain one event with pictures and numbers

For ID 3, the recorded center moves from y = 411.236 in frame 36 to y = 414.239 in frame 37. The rule accepts frame 37 because 411.236 < 413 <= 414.239. Its current x is approximately 474.9, inside the permitted span. ID 3 has not yet been counted.

The same ID before, at and after its accepted event. Coordinates are rounded for display; the rule uses the original values.

This evidence is more informative than a final total alone. You can inspect the visible vehicle, the relevant center positions and the exact decision together.

Equal totals can hide different events

Move only the line upward by 30 pixels, from y = 413 to y = 383. Reuse the same recorded tracks and horizontal span. The result is again 7 events, but the accepted IDs differ:

Line	Events	Accepted IDs
y = 413	7	4, 3, 2, 5, 9, 7, 55
y = 383	7	3, 2, 5, 9, 7, 55, 48

ID 4 is first observed at y ≈ 398.5, already below the shifted line. Its observations support a crossing of y = 413, but not y = 383. ID 48 supplies an event only at the shifted line. Comparing event membership exposes what the matching totals hide.

Check logic separately from tracking quality

Run these checks after defining count_passes above. Their deliberately chosen center positions isolate Python logic from model quality. The three assertions test repeated downward crossings by one ID, an ID first seen below the line, and upward movement. All x positions are inside the counting segment.

def sample(y_values):
    return [{"frame_idx": i, "objects": [{"track_id": 1,
             "class_name": "car", "center": (200, y)}]}
            for i, y in enumerate(y_values)]

assert len(count_passes(sample([411, 413, 410, 416]), 413, (80, 1000))) == 1
assert len(count_passes(sample([414, 420]), 413, (80, 1000))) == 0
assert len(count_passes(sample([420, 410]), 413, (80, 1000))) == 0
print("All three counting-rule checks passed.")

The expected counts are 1, 0 and 0. After these checks pass, inspect a short video interval manually and match individual passages to event rows.

This compact rule has limits. A changed ID can allow the same vehicle to count again. Missing observations may leave a long gap between the two positions used by the comparison. There is no gap limit or hysteresis band here, and motion upward is intentionally excluded. Real deployment needs representative footage and separate checks for missed passages and false counts.

For the tracking API and tracker configuration, see the official Ultralytics tracking documentation. This example explicitly selects bytetrack.yaml; it does not rely on whichever tracker a library version uses by default.

If you want to build the full project, including the notebook, local application and practical exercises, see my English course: Computer Vision: OpenCV, Machine Learning and YOLO.

New Udemy Course: Computer Vision with OpenCV, Machine Learning and YOLO

I have published my new English-language course on Udemy: Computer Vision: OpenCV, Machine Learning and YOLO.

The course takes you from working with image arrays to building Python applications that detect, track and count objects in images and video. It combines theory with separate practical lessons, so you can see the inputs, follow the code and inspect the results.

What you will learn

• Python, NumPy and OpenCV: image representation, color channels, transformations, filtering and edge detection.
• Machine learning: feature extraction, classification, model comparison and evaluation with scikit-learn.
• Deep learning and CNNs: neural network foundations, convolution, pooling, image classification and transfer learning with TensorFlow and PyTorch.
• YOLO: object detection, custom training, object tracking, segmentation and image classification.
• Pose estimation and deployment: body keypoints, movement features, model export, CPU performance measurements and comparisons with alternative detectors.

Four projects that bring the ideas together

• Parking Lot Monitor: work with parking-space regions and occupancy information.
• Traffic Counter: detect and track vehicles, then count crossings of a defined line.
• Fitness Tracker: use pose keypoints and movement states to count squat repetitions.
• Document Scanner: find document boundaries and correct perspective.

These are learning projects that make the connection between model outputs and application logic visible.

Video lessons, notebooks and knowledge checks

The course includes more than 10 hours of English video, downloadable notebooks and lesson resources, and 12 section quizzes with 36 questions. Practical lessons include setup and dependency instructions to help you run the examples locally.

It is aimed at developers, students and engineers who want to move from computer vision concepts to working Python examples. Some programming experience is helpful; the opening sections review Python and NumPy foundations.

Launch offer: US$12.99

Coupon code: CBEE0F25D4870B5F9123

Open the course with the US$12.99 coupon

The coupon is valid until 27 November 2026 at 06:20 UTC (26 November, 22:20 PST). Check that the coupon is applied and review the final price shown by Udemy before purchasing.

You can also use my standard Udemy course link to view the course and its curriculum. This separate link does not include the coupon code.

I would be glad to hear which computer vision applications you would like to build.