Computer-Vision_Soccer

Using Explainability Methods for an Object Detection Modell from Roboflow.

Research Question: Which visual aspects influence the model's decision in ball detection?

Inspired by: https://github.com/roboflow/sports

Dataset: https://universe.roboflow.com/roboflow-jvuqo/football-ball-detection-rejhg/

Modell: YOLOv8

Lime Analysis:

• Blurring different pixel areas to identify which changes influence the model's decision (model-agnostic).
• Inspired by: https://github.com/marcotcr/lime

SHAP Analysis

• Theorie: SHAPley-Values: Determining which pixels contribute the most to the model's decision.

Grad-CAM Analysis

• I had to train the YOLOv8 because the weights are only available in the roboflow premium version
• The required weights are in the folder models/football_gradcam/weights so the training can be skipped

Setup und Execution

• python environment: >= 3.8 (and <3.13)
• Roboflow: API-Key needed
• for Configuration: see config.py
• UNZIP needed for the download of the data set
• Execution should be performed from the project root directory

Steps:

• run the first cells of setup_model.ipynb
  • Installing Requirements
  • Setup of the Model
  • Download of the Dataset
• Now you're free to run the other notebooks, from the project root