GNN

Relational Group Activity Recognition Table of Contents Key Updates Introduction How the Relational Layer Works Usage Clone the Repository Install Dependencies Download Model Checkpoint Option 1: Use Python Code Option 2: Download Directly Dataset Overview Example Annotations Train-Test Split Dataset Statistics Group Activity Labels Player Action Labels Dataset Organization Dataset Download Instructions Ablation Study Baselines Single Frame Models Performance Comparison (Original Paper) My Scores (Accuracy) Temporal Models Performance Comparison (Original Paper) My Scores (Accuracy) Attention Models (New Baseline) My Scores (Accuracy) Confusion Matrix Key Updates ResNet-50 Backbone: Replaced VGG19 with ResNet-50 for stronger feature extraction. Ablation Studies: Comprehensive experiments to evaluate the contribution of each model component. Test-Time Augmentation (TTA): Implemented to improve robustness and reliability during inference. Graph Attention Operator: Implementation for an attention-based relational layer. Improved Performance: Achieves consistently higher accuracy across all baselines compared to the original paper. Modern Implementation: Fully implemented in PyTorch with support from PyTorch Geometric. Introduction Traditional pooling methods (max, average, or attention pooling) reduce dimensionality but often discard important spatial and relational details between people. The Hierarchical Relational Network (HRN) addresses this by introducing a relational layer that explicitly models interactions between individuals in a structured relationship graph. ...