Abstract
Dynamic point cloud videos exhibit intricate spatio-temporal coupling. Current end-to-end methods are only evaluated on specific datasets and tasks, generalize poorly across scenes or tasks, and must be retrained from scratch. To uncover transferable universal priors, we refocus our investigation on static point cloud foundation models. In this paper, we first identify two bottlenecks in existing transfer learning methods: Heavy computation and redundant design. To obtain a compact and lightweight solution, we propose an efficient cross-modal transfer method that uses progressive adapters. It reduces self-attention from quadratic to linear complexity with non-overlapping sliding window attention. A shuffled Hilbert-Z bidirectional scan curve converts the dynamic point cloud video into a Mamba-compatible feature sequence; progressive gating fusion then lets the foundation model and the newly added adapter cooperate efficiently. The foundation model weights remain unchanged and no external auxiliary designs are required. Experiments on MSR-Action3D, HOI4D and SHREC'17 show that tuning only 1.8% of the parameters already surpasses existing baselines, confirming the superiority of our method.
| Translated title of the contribution | Efficient Transfer Learning for Dynamic Point Clouds With Progressive Adapters |
|---|---|
| Original language | Chinese (Traditional) |
| Pages (from-to) | 1710-1720 |
| Number of pages | 11 |
| Journal | Zidonghua Xuebao/Acta Automatica Sinica |
| Volume | 52 |
| Issue number | 8 |
| DOIs | |
| State | Published - Aug 2026 |
| Externally published | Yes |
Fingerprint
Dive into the research topics of 'Efficient Transfer Learning for Dynamic Point Clouds With Progressive Adapters'. Together they form a unique fingerprint.Cite this
- APA
- Author
- BIBTEX
- Harvard
- Standard
- RIS
- Vancouver