IIIT Hyderabad

Skeletal Human Activity Recognition

• Recognized human activities from skeleton representations using Hierarchical Deep Learning models
• Studied architectures like Hierarchical Deep CNN, Network of Experts, Deep Mixture Of Experts via Shallow Embedding, to improve validation accuracy on NTU 120 data-set by clustering similar classes
• Implemented Network of Experts architecture, a hierarchical deep learning architecture in PyTorch
• Fine-tuned model along with 5 experts with overall soft-max loss yielded a validation accuracy of 84.28%