Temporal Sequence Modeling for Sports Action Recognition

This project focuses on fine-grained sports action recognition using two main architectures:

1. CNN-based Sequence Models: These models combine CNNs for feature extraction with RNNs(GRU layers) for temporal sequence modeling:

   • VGG19
   • InceptionV3
   • InceptionV4-ResNet (hybrid model)
   • EfficientNetB4

2. ViViT (Video Vision Transformer): A pure transformer-based approach for end-to-end video classification, capturing both spatial and temporal features.

Model Architectures

1. CNN-based Sequence Models

• Feature Extractors: VGG19, InceptionV3, InceptionV4-ResNet, EfficientNetB4
• Temporal Model: GRU layers

2. ViViT Model

• Transformer-based model for video classification
• Spatiotemporal attention and tubelet embedding

Evaluation

Each model is evaluated using:

• Accuracy, Precision, Recall, F1-Score
• Training/validation curves
• Confusion matrix

Acknowledgments

• Dr. Lina Chato
• UCF101 dataset
• TensorFlow team
• All the cited authors