Learning multi-granularity skeleton representations via hierarchical graph contrastive learning
Yaxiong Liu, Xianfeng Zhai · 2026
Read Learning multi-granularity skeleton representations via hierarchical graph contrastive learning by Yaxiong Liu, Xianfeng Zhai (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/pmc-13483196.txt.
Description
Liu and Zhai proposed a self-teaching method for recognizing human actions from skeleton data, the stick-figure joint tracks that cameras extract from video. Labels for training are scarce, and standard tricks that warp the data can wreck the skeleton's structure. Their 2026 method, called HGCL, learns at two scales at once, whole-body motion and single body parts, while a contrast task teaches it which clips match.
On two public video benchmarks the method reached about 91 percent accuracy on the larger set and about 67 percent on the harder one in the reported tests. Its data warping respects bone lengths, with a proven cap on how far bones can distort. Detection confidence for each joint flows through the whole pipeline, muting unreliable joints instead of trusting them. Liu argues coupling the scales, rather than bolting on separate modules, is what makes the learning work.
Catalog record
| CREATOR | Yaxiong Liu, Xianfeng Zhai |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 116 KB |
| VERSION | 1.0 |
| LICENSE | CC BY 4.0 |
| RIGHTS | CC BY |
| SOURCE | Europe PMC |