Thread by @cgmsnoek

ACM #Multimedia 2021: Skeleton-Contrastive 3D Action Representation Learning w/ @fmthoker, @doughty_hazel: arxiv.org/abs/2108.03656 We learn invariances to multiple #skeleton representations and introduce various skeleton augmentations via noise contrastive estimation 1/n

Contribution I: leverage multiple input-representations of #3D-#skeleton sequences. Our inter-skeleton contrast learns from a pair of representations in a cross-contrastive fashion. Enriches the sparse input space and focuses on the high-level semantics of the skeleton data. 2/n

Contribution II: three skeleton-specific #augmentations for generating positive pairs which encourage the model to focus on the spatio- temporal #dynamics of skeleton-based @action sequences, ignoring confounding factors such as viewpoint and exact joint positions. 3/n

Our approach achieves state-of-the-art performance for self-supervised learning from #skeleton data on the challenging #PKU and #NTU datasets with multiple downstream tasks, including #action #recognition, #action #retrieval and #semi-#supervised #learning. 4/n

Code is here: github.com/fmthoker/skele… n/n