• Overview
• Installation Guide
• Usage
• Files in repo
• References

In this project we used fine-tuning of a YOLOv7 network and implementation and training of a BYOL network to classify, localise and split into teams, players and other objects in football games.

Draws inspiration from a similar basketball-based implementation by James Skelton - https://blog.paperspace.com/yolov7/

In this notebook, we will explain and implement the algorithm.

We can divide the workflow into 2 main steps:

1. Localisation and Classification of football match image or video.
2. Clustering of players from image to their two teams.

Combining the results from these steps should yield the desired result. The workflow can be depicted as follows:

(for example)

python3 -m venv venv
source venv/bin/activate

pip install -r requirements.txt

Download the weights for each of the networks, they're stored here: https://drive.google.com/drive/folders/1rpshi_ZycbRSJtwy1QXx73oTUWTzcJgk?usp=sharing

Open a terminal inside the cloned repo and run the following:

python run.py --yolo-weights 'path/to/YOLO/weights.pt' --byol-weights 'path/to/BYOL/weights.pt' --source 'path/to/source/image_or_video'

• yolo-weights = pretrained weights for YOLO network
• byol-weights = pretrained weights for BYOL network
• source = path to source image or video
• img-size = inference size in pixels (default=640)
• conf-thres = object confidence threshold (default=0.25)
• iou-thres = IOU threshold for NMS (default=0.45)
• device = cuda device, i.e. 0 or 0,1,2,3 or cpu (default='')
• view-img = display results
• save-txt = save results to *.txt
• save-conf = save confidences in --save-txt labels
• nosave = do not save images/videos
• classes = filter by class: --class 0, or --class 0 2 3 (0:ball, 1:goalkeeper, 2:player, 3:referee)
• agnostic-nms = class-agnostic NMS
• augment = augmented inference
• project = save results to project/name (default='detections')
• name = save results to project/name (default='run')
• exist-ok = existing project/name ok, do not increment
• no-trace = don`t trace model

[1] Official YOLOv7 - Implementation of paper - https://github.com/WongKinYiu/yolov7

[2] Paperspace Blog "How to train and use a custom YOLOv7 model" - https://blog.paperspace.com/yolov7/

[3] BYOL - Technion DL Tutorial - Tal Daniel - https://github.com/taldatech/ee046211-deep-learning/blob/main/ee046211_tutorial_09_self_supervised_representation_learning.ipynb