SRENet: Spectral Re-Entry Network for Point Cloud Action Recognition

Researchers have proposed multiple frameworks for action recognition, including SRENet, a spectral-aware approach for point cloud sequences, and SkelHCC, a hyperbolic CLIP-driven cache adaptation method for one-shot skeleton-based recognition.

SRENet, introduced in a paper on arXiv^[1], is designed to learn global context and fine-grained temporal dynamics from a frequency perspective. It features a Spectral Decomposition Block (SDeBlock) and a Spectral Re-entry Block (SReBlock) to analyze point cloud sequences. SRENet achieved state-of-the-art performance on MSR-Action3D, NTU-RGBD, and NTU-RGBD120 datasets. Another approach, proposed in a separate arXiv paper^[2], uses sparse point trajectories and a simple transformer architecture for 2.5D trajectory-based recognition, achieving 45% top-1 accuracy on Something-Something V2 and 54% on EPIC-Kitchens-100^[2]. SkelHCC, presented in a third arXiv paper^[3], addresses one-shot action recognition using a hyperbolic CLIP-driven cache adaptation approach, outperforming state-of-the-art methods on NTU RGB+D 60, NTU RGB+D 120, and PKU-MMD datasets. The SRENet paper was submitted on June 2, 2026^[1][3].