Gait attributes get a 242,729-sequence benchmark, and its pretrained encoder swaps into identity recognition
Synopsis
Gait attribute recognition now has a large multimodal benchmark: RA-GAR covers 1,001 subjects and 242,729 walking sequences annotated with 15 static and motion gait attributes, addressing the shortage of fine-grained annotation. The accompanying Gait-VL matches attribute-specific multi-frame visual evidence against textual class prototypes instead of representing candidate classes through learned classifier output weights. It reaches state-of-the-art attribute recognition on RA-GAR, and replacing generic upstream vision models with the attribute-pretrained encoder yields consistent gains across several identity recognition pipelines and benchmarks.
Interpretation
Gait attribute recognition now has a large multimodal benchmark with fine-grained annotation, covering 1,001 subjects, 242,729 walking sequences and 15 static and motion gait attributes. Earlier benchmarks supplied only limited fine-grained annotation, leaving attribute recognition short of supervision. The benchmark reports its scale in subjects, walking sequences and attribute classes, and attribute recognition experiments on it reach state-of-the-art performance.
Recognition shifts from comparing classifier weights to directly matching visual evidence with textual class prototypes. Classifier-based methods represented candidate classes through learned output weights rather than aligning visual evidence with textual class representations. Gait-VL achieves state-of-the-art attribute recognition performance on RA-GAR.
The visual encoder obtained from attribute pretraining can replace generic upstream large vision models in identity recognition pipelines. Attribute pretraining uses no identity classification or metric-learning objectives, so the swap tests the transferability of the attribute representation itself. Consistent improvements appear across multiple identity recognition pipelines and benchmarks after the replacement.
Perspective
The work targets research and engineering settings in gait attribute recognition and identity recognition: RA-GAR provides annotation for 1,001 subjects, 242,729 walking sequences and 15 static and motion gait attributes, and Gait-VL performs attribute recognition on that annotation. The transferability claim is limited to replacing the attribute-pretrained visual encoder inside identity recognition pipelines, validated across multiple pipelines and benchmarks.
Attribute pretraining uses no identity classification or metric-learning objectives, so its transfer gains come from the attribute representation itself; the material is at summary level, and the specific numbers, pipeline configurations and benchmark names for attribute recognition and downstream identity recognition are not expanded in the visible scope.