[ACM MM'25] Seeing from Magic Mirror: Contrastive Learning from Reconstruction for Pose-based Gait Recognition
Official repository of MirrorGait code and the StreamGait dataset.
- StreamGait: A large-scale, unlabelled gait dataset collected from 353 YouTube live-stream channels across 30 countries / 81 cities.
- MirrorGait: A self-supervised 3D-aware pre-training framework that lifts 2D poses to 3D for contrastive learning, achieving SOTA on Gait3D, GREW and OUMVLP-Pose with minimal fine-tuning.
- Release the data acquisition procedures and protocols.
- Release the training code and models of MirrorGait.
| Attribute | StreamGait | GaitLU-1M |
|---|---|---|
| #Sequences | 854 K | 1.0 M |
| Camera | Real-world fixed | Handheld |
| Environments | 81 cities, 30 countries | 16 cities |
| Benchmark | Rank-1 (prior SOTA) | Rank-1 (MirrorGait-ft) | Gain |
|---|---|---|---|
| Gait3D | 38.1 (SkeletonGait) | 48.1 | +10.0% |
| GREW | 77.4 (SkeletonGait) | 79.6 | +2.2% |
| OUMVLP-Pose | 70.5 (GaitGraph2) | 73.9 | +3.4% |
For questions or collaboration: Shibei Meng, mengshibei@mail.bnu.edu.cn