Introduction
This repository is an implementation of "Exploiting deep residual networks for human action recognition from skeletal data" by Huy-Hieu Pham, Louahdi Khoudour, Alain Crouzil, Pablo Zegers, and Sergio A. Velastin. You can train deep Residual Networks on skeletal data for action recognition tasks from scratch.
Training from scratch
Before running the experiments, you need to download and compile VLFeat, and MatConvNet. Then, start to train ResNets on GPU by the following commande:
experiments([20 32 44 56 110], 'resnet', 'gpus', [1]);
Experimental results
Learning curves on KARD dataset. Dashed lines denote training errors (%), bold lines denote test errors (%).
Learning curves on MSR Action3D dataset. Dashed lines denote training errors (%), bold lines denote test errors (%).
Learning curves on NTU-RGB+D dataset. Dashed lines denote training errors (%), bold lines denote test errors (%).