Research map / Label-efficient and robust learning / Recognition and long-tail
WaveLoss: An Adaptive Dynamic Loss for Deep Gait Recognition
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes WaveLoss, a part-level loss for gait recognition that interpolates between pre- and post-fusion via norm-fusion, with a sinusoidally varying exponent (dynamic learning process), gradient truncation and normalization.
Problem
Part-level loss fusion in deep gait recognition: post-fusion is sensitive to noisy parts while pre-fusion gets trapped in zero-gradient regions late in training; existing gait methods mostly use triplet and softmax losses rather than designed metric losses.
Approach
Norm-fusion weights each part's loss by an exponentiated, l2-normalized difficulty term raised to an exponent l that interpolates between post-fusion (l=0) and pre-fusion-like behavior; l is varied sinusoidally during training, and gradient truncation and normalization stabilize optimization.
Key results
- On Gait3D, WaveLoss reaches 75.6% Rank-1 and 66.5% mAP, above DeepGaitv2 (74.4% Rank-1).
- On CASIA-B, WaveLoss averages 94.7% across NM/BG/CL, versus 94.2% for CircleLoss post-fusion and 80.4% for pre-fusion.
- On OU-MVLP, WaveLoss achieves 89.7% mean Rank-1, a 0.2% gain over DeepGaitV2 baseline per the text (DeepGaitV2 reported 91.9 in table).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use Waveloss instead of the baseline loss | DeepGaitV2 · OU-MVLP | Rank-1 accuracy: 91.9 -> 92.1 | Helped |
| Use Waveloss instead of the baseline loss | GaitGL · CASIA-B mean | Rank-1 accuracy | Helped |
| Use Waveloss instead of DeepGaitV2's loss | DeepGaitV2 · Gait3D | Rank-1 accuracy: 74.4 -> 75.6 | Helped |
| Use a dynamic l range [-1,1] rather than static l=-1 (ablation) | Waveloss on GaitGL · CASIA-B mean | Rank-1 accuracy: 94.4 -> 94.7 | Helped |
| Use a dynamic l range [-1,1] rather than static l=0 (ablation) | Waveloss on GaitGL · CASIA-B mean | Rank-1 accuracy: 94.2 -> 94.7 | Helped |
| Use a dynamic l range [-1,1] rather than static l=1 (ablation) | Waveloss on GaitGL · CASIA-B mean | Rank-1 accuracy: 94.4 -> 94.7 | Helped |
| Replace norm-fusion with pre-fusion (ablation) | Waveloss on GaitGL · CASIA-B mean | Rank-1 accuracy: 94.7 -> 80.4 | Hurt |
| Replace norm-fusion with post-fusion (ablation) | Waveloss on GaitGL · CASIA-B mean | Rank-1 accuracy: 94.7 -> 94.2 | Hurt |
| Use norm-fusion with CircleLoss (ablation) | GaitGL with CircleLoss · CASIA-B mean | Rank-1 accuracy: 94.2 -> 94.4 | Helped |
| Add Dynamic Learning Process to CircleLoss (ablation) | GaitGL with CircleLoss · CASIA-B mean | Rank-1 accuracy: 94.2 -> 94.4 | Helped |
| Use Waveloss (norm-fusion and Dynamic Learning Process) instead of CircleLoss (ablation) | GaitGL with CircleLoss · CASIA-B mean | Rank-1 accuracy: 94.2 -> 94.7 | Helped |
Surprising findings
- Pre-fusion performs far worse than post-fusion on CASIA-B (80.4% vs 94.2% mean), with CL accuracy dropping to 63.7%.
- Fixed exponent values (l = -1, 0, 1) give nearly the same accuracy as each other, with the dynamic schedule giving only a small further gain (94.4-94.2 vs 94.7).
Limitations
- Gains over prior methods are small (often 0.2% to 0.5% on CASIA-B and OU-MVLP).
- Hyperparameters such as l range, number of cycles C, and sensitivity are only partially explored.
- Evaluated only on gait datasets with GaitGL and DeepGaitV2 backbones.
Experimental setup
Datasets: CASIA-B OU-MVLP Gait3D
Benchmarks: CASIA-B OU-MVLP Gait3D
Metrics: Rank-1 accuracy Rank-5 accuracy mAP mINP mean accuracy across views
Baselines: GaitSet GaitGL DANet DeepGaitv2 GaitMPL DyGait SMPLGait Softmax Loss Triplet Loss Circleloss
Builds on: CircleLoss GaitGL DeepGaitV2 ISR OpenGait GaitMPL
Closest papers by meaning
- Learning Visual Prompt for Gait Recognition (CVPR 2024 · similarity 0.84)
- Unlocking Motion from Large Vision Models with a Semantic and Kinematic Duality for Gait Recognition (CVPR 2026 · similarity 0.83)
- BiggerGait: Unlocking Gait Recognition with Layer-wise Representations from Large Vision Models (NeurIPS 2025 · similarity 0.83)
- On Denoising Walking Videos for Gait Recognition (CVPR 2025 · similarity 0.82)
- RA-GAR: A Richly Annotated Benchmark for Gait Attribute Recognition (AAAI 2025 · similarity 0.82)
- QAGait: Revisit Gait Recognition from a Quality Perspective (AAAI 2024 · similarity 0.81)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.