Pose-Based Person Identification
Deep learning for identity from body pose
Overview
Compared feedforward and bidirectional GRU architectures for identifying individuals from body pose sequences using the PoseTrack21 dataset. Automated hyperparameter tuning with Optuna on NVIDIA H100 GPU nodes via the University of Sheffield HPC cluster. Augmented training sequences using a sliding window approach and evaluated models with top-1 accuracy, calibration metrics, and occlusion-based ablation tests.
Key features
- Compared feedforward vs bidirectional GRU architectures for sequence-based person identification
- Automated hyperparameter search with Optuna across HPC GPU nodes
- Sliding window augmentation for training sequence generation
- Occlusion-based ablation tests across body joints to identify discriminative keypoints
- Comprehensive evaluation: top-1 accuracy, Expected Calibration Error, reliability diagrams, confusion matrices
Approach
Pose sequences from PoseTrack21 were extracted and preprocessed into fixed-length windows using a sliding window approach. Two model architectures were built in PyTorch: a feedforward network and a bidirectional GRU. Hyperparameters (layer sizes, learning rate, dropout, etc.) were optimized using Optuna. Models were evaluated on held-out test sequences with metrics including top-1 accuracy, calibration error, and visual analysis via reliability diagrams and confusion matrices. Ablation studies systematically occluded individual body joints to measure their contribution to identification accuracy.
Results
Successfully demonstrated that bidirectional GRU architectures outperform feedforward networks for pose-based person identification. The ablation studies revealed which body joints are most discriminative for identity. All visualisations were generated with Matplotlib and Seaborn.