SkateFormer: Skeletal-Temporal Transformer for Human Action Recognition

Jeonghyeok Do     Munchurl Kim†
†Corresponding author
Korea Advanced Institute of Science and Technology (KAIST), South Korea      
ECCV 2024
Paper arXiv Video Code

Overview of SkateFormer's partition-specific attention strategy. SkateFormer partitions joints and frames based on different types of skeletal-temporal relation (4 Skate-Types) and performs skeletal-temporal self-attention (Skate-MSA) within each partition.

Abstract

We propose a Skeletal-Temporal Transformer (SkateFormer), a partition-specific attention strategy (Skate-MSA) for skeleton-based action recognition that captures skeletal-temporal relations and reduces computational complexity.

We introduce a range of augmentation techniques and an effective positional embedding method, named Skate-Embedding, which combines skeletal and temporal features. This method significantly enhances action recognition performance by forming an outer product between learnable skeletal features and fixed temporal index features.

Our SkateFormer sets a new state-of-the-art for action recognition performance across multiple modalities (4-ensemble condition) and single modalities (joint, bone, joint motion, bone motion), showing notable improvement over the most recent state-of-the-art methods. Additionally, it concurrently establishes a new state-of-the-art in interaction recognition, a sub-field of action recognition.

Network Architecture


Top. The overall framework of SkateFormer.
Bottom left. The Skate-MSA of SkateFormer.
Bottom right. Skate-Type partition and reverse.

Discussion on Various Partition-Based Approaches


Comparison with existing transformer-based methods. S-Type, T-Type, S-Attn, T-Attn, T-Conv, and ST-Attn indicate 'physically neighboring joints', 'local motion', ‘skeletal attention’, ‘temporal attention’, ‘temporal convolution’ and ‘skeletal-temporal attention’, respectively.

Ensemble Performance Comparison


Top-1 accuracy of skeleton-based action recognition methods. (i) E1- joint modality only; (ii) E2 - joint + bone modalities; and (iii) E4 - joint + bone + joint motion + bone motion modalities.


Top-1 accuracy of skeleton-based interaction recognition methods. Human interaction recognition is a sub-part of skeleton-based action recognition, specifically focusing on scenarios where two or more individuals coexist within a single action.

Computational Complexity Comparision


Comparative analysis of SkateFormer with other methods by parameters, FLOPs, inference time, and average top-1 accuracy for joint modality.

Single Modality Comparision


(i) J - joint modality; (ii) B - bone modality; (iii) JM - joint motion modality; and (iv) BM - bone motion modality.

Analysis of Partition-Specific Attention


The activation level of Skate-Types according to action labels. For 'sitting down' (action label: 7) and 'standing up' (action label: 8), Skate-Type-3 or Skate-Type-4 exhibited pronounced activations, while 'reading' (action label: 10) and 'writing' (action label: 11) prominently activated Skate-Type-1.

BibTeX

@inproceedings{do2025skateformer,
  title={Skateformer: skeletal-temporal transformer for human action recognition},
  author={Do, Jeonghyeok and Kim, Munchurl},
  booktitle={European Conference on Computer Vision},
  pages={401--420},
  year={2025},
  organization={Springer}
}