BROWSE
Detail
Journal Article
Long-Tailed Skeleton-Based Action Recognition via Realistic Mixing and Class-Aware Sampling
- Authors
- Jinwoo Kim, Wonhee Kim, Mi-Seon Kang, Jong Taek Lee
- Issue Date
- 2026-04
- Citation
- IEEE Access, v.14, pp.52680-52689
- ISSN
- 2169-3536
- Publisher
- IEEE
- Language
- English
- Type
- Journal Article
- Abstract
- Skeleton-based action recognition has gained increasing attention due to its robustness to background variations and privacy-preserving properties. However, real-world skeleton datasets often exhibit severe class imbalance, leading to long-tailed data distributions that significantly degrade recognition performance, especially for tail classes. Existing mix-based skeleton augmentation methods partially alleviate this issue by increasing data diversity, but they often generate physically implausible skeletons and fail to explicitly account for class imbalance during augmentation. In this paper, we propose a long-tailed augmentation framework with two key components: 1) realistic skeleton mixing in the relative bone vector space and 2) class-aware sampling that increases effective augmentation frequency of underrepresented classes. Specifically, we perform mixing in the relative bone vector space rather than joint coordinate space to preserve kinematic consistency and further enforce bone-wise length alignment to prevent unrealistic limb deformations. To mitigate class imbalance, we introduce a class-aware weighting mechanism that biases sampling toward tail classes. Extensive experiments on the long-tailed NTU RGB+D and NTU RGB+D 120 benchmarks show that the proposed method consistently outperforms existing approaches across various imbalance factors, delivering notable improvements on the tail classes while remaining robust across different backbone architectures.
- Keyword
- Augmentation, deep learning, long-tailed distribution learning, skeleton-based action recognition
- KSP Keywords
- Action recognition, Data Distribution, Distribution learning, Existing Approaches, Key Components, Long-tailed distributions, Privacy-preserving, Real-world, Recognition performance, class imbalance, deep learning(DL)
This work is distributed under the term of Creative Commons License (CCL)
(CC BY)
(CC BY)
