Unsupervised Feature Learning of Human Actions as Trajectories in Pose Embedding Manifold

• DOI: 10.48550/arxiv.1812.02592
• Publication Date: 2018-01-01
• Authors: Jogendra Nath Kundu; Maharshi Gor; Phani Krishna Uppala; R. Venkatesh Babu
• Keywords: Feature Learning; Autoencoder; Representation; Feature (linguistics); Sequence (biology); Feature vector; Manifold (fluid mechanics)

An unsupervised human action modeling framework can provide useful pose-sequence representation, which be utilized in a variety of pose analysis applications. In this work we propose novel temporal framework, embed the dynamics 3D human-skeleton joints to continuous latent space an efficient manner. contrast end-to-end explored by previous works, disentangle task individual representation learning from actions as trajectory embedding space. order realize manifold with improved reconstructions, unsupervised, procedure named Encoder GAN, (or EnGAN). Further, use embeddings generated EnGAN model using bidirectional RNN auto-encoder architecture, PoseRNN. We introduce first-order gradient loss explicitly enforce regularity predicted motion sequence. A hierarchical feature fusion technique is also investigated for simultaneous local skeleton along global variations. demonstrate state-of-the-art transfer-ability learned against other supervisedly and unsupervisedly fine-grained recognition on SBU interaction dataset. show qualitative strengths proposed visualizing reconstructions interpolations pose-embedding space, low dimensional principal component projections reconstructed trajectories.