New Implementation of 'A Hierarchical Deep Temporal Model for Group Activity Recognition' (CVPR 2016)

Re-implemented and enhanced a Hierarchical Deep Temporal Model for group activity recognition

(CVPR 2016), with significant performance improvements.

● Developed and evaluated multiple baselines:

○ Frame-level classifier (ResNet-50): 85.1% accuracy.

○ Person-level features (ResNet-50 + pooling): 75.2% accuracy.

○ Temporal model (ResNet-50 + LSTM, 9-frame sequences): 86.6% accuracy.

● Applied on a volleyball dataset with both frame-level and person-level annotations.

● Tech stack: PyTorch, Torchvision, ResNet-50, LSTM, Scikit-learn.