VolleyballObjectDetection

1. Description

This repository contains the code for the project of the Polytechnic Uniersity of Turin course Image Processing and Computer Vision; the project consists in the developement of object detection models (YOLO-based) to detect players, court lines and the ball in a volleyball amateur match video stream. The pipeline supports:

• robustness to lighting and color variations
• automatic zoom on the ball during actions

More details in the project report.

2. Project structure

• scripts/ — inference, preprocessing and training utilities (notably scripts/detect_players.py).
• datasets/ — YOLO-formatted datasets (images, labels, data.yaml).
• frames/ — extracted and cropped frames for inspection.
• models/ — bundled model weights (example: yolov8m.pt, best.pt).
• runs/ — training run outputs (weights, CSV metrics, plots, sample predictions).
• predictions/ — saved inference outputs on test frames.
• params/ — experiment parameter files and prediction summaries.

3. Dependencies

• Python 3.10+
• Install with conda (recommended for torch/torchvision to ensure compatibility with CUDA):

conda create -n volley python=3.10 -y
conda activate volley
# install torch via conda (choose CUDA-enabled package if appropriate)
conda install pytorch torchvision -c pytorch -y

• Install the remaining Python packages with pip:

pip install -r requirements.txt

4. How to visualize results

Use scripts/detect_players.py to open a test clip and visualize detections. Example (with the conda env active):

python scripts/detect_players.py --clip_number 1

• --clip_number selects test_clip_<n>.mp4 at the repository root (valid values: 1 or 2).
• The viewer window is interactive; press q to quit.
• Model paths are configured via constants at the top of scripts/detect_players.py — update them to point to different weights if needed.

For further analysis, inspect runs/<run_name>/ (weights, results.csv, plots) and the predictions/ folder for example annotated frames.