Research map / Video / Video segmentation and tracking
GaitCycFormer: Leveraging Gait Cycles and Transformers for Gait Emotion 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 GaitCycFormer, a graph-transformer for skeleton-based gait emotion recognition that uses a window-based gait cycle detector, cycle position encoding, and a bi-level intra/inter-cycle Transformer on top of ST-GCN.
Problem
Skeleton-based gait emotion recognition struggles to learn long-range temporal features and is sensitive to the order of sampled frames, hurting accuracy under random sampling.
Approach
Segments gait sequences into gait cycles with a window-based detector and encodes each frame's position within its cycle (CPE); a bi-level Transformer applies attention within cycles (intra-cycle) and across cycles at matching positions (inter-cycle) inside GTT blocks that combine GCN and Transformer.
Key results
- Achieves 86.3% accuracy on EGait, above STA-GCN's 85.8% and the prior best reported in Table 1.
- Achieves 83.4% accuracy on EWalk, above STA-GCN's 82.8%.
- Under random sampling with 48 frames, accuracy drops only 0.1% (83.7% to 83.6%), while TCN drops 6.4% (78.2% to 71.8%).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use GaitCycFormer instead of the strongest listed baseline, STA-GCN | EGait · EGait | accuracy: 85.8% -> 86.3% | Helped |
| use GaitCycFormer instead of STA-GCN | EWalk · EWalk | accuracy: 82.8% -> 83.4% | Helped |
| add GTT blocks to the baseline (ablation) | GaitCycFormer on EGait · EGait | accuracy: 79.8% -> 86.3% | Helped |
| add the Inter-cycle Transformer (ablation) | GaitCycFormer on EGait · EGait | accuracy: 83.6% -> 86.3% | Helped |
| add the Intra-cycle Transformer (ablation) | GaitCycFormer on EGait · EGait | accuracy: 83.2% -> 86.3% | Helped |
| add positional encoding (ablation) | Vanilla Transformer on EGait · EGait | accuracy: 82.6% -> 83.2% | Helped |
| replace vanilla positional encoding with Cycle Position Encoding (ablation) | Vanilla Transformer on EGait · EGait | accuracy: 83.2% -> 84.6% | Helped |
| add positional encoding (ablation) | Intra-Inter Transformer on EGait · EGait | accuracy: 84.5% -> 85.3% | Helped |
| replace vanilla positional encoding with Cycle Position Encoding (ablation) | Intra-Inter Transformer on EGait · EGait | accuracy: 85.3% -> 86.3% | Helped |
| use Intra-Inter Transformer instead of Vanilla Transformer with CPE (ablation) | EGait · EGait | accuracy: 84.6% -> 86.3% | Helped |
| randomly sample 120 rather than regularly sample 120 frames (ablation) | TCN on EGait · EGait, 120 frames | accuracy: 78.8% -> 74.5% | Hurt |
| randomly sample 48 rather than regularly sample 48 frames (ablation) | TCN on EGait · EGait, 48 frames | accuracy: 78.2% -> 71.8% | Hurt |
Surprising findings
- The vanilla TCN baseline loses 6.4 points under random 48-frame sampling, while the proposed model is nearly unchanged.
Limitations
- Gait cycle detection relies on heuristic foot-contact rules and a hyperparameter a=15 set by frame rate.
- Only evaluated on two small gait emotion datasets (EGait, EWalk).
Experimental setup
Datasets: EGait EWalk
Benchmarks: EGait EWalk
Metrics: top-1 accuracy
Baselines: STEP TT-GCN G-GCSN MSA-GCN STA-GCN ProxEmo TNTC CAGE SSAL TCN
Builds on: ST-GCN Transformer Graph Transformer over-globalizing analysis (Xing et al. 2024)
Closest papers by meaning
- RA-GAR: A Richly Annotated Benchmark for Gait Attribute Recognition (AAAI 2025 · similarity 0.84)
- SkeletonGait: Gait Recognition Using Skeleton Maps (AAAI 2024 · similarity 0.83)
- Bridging Gait Recognition and Large Language Models Sequence Modeling (CVPR 2025 · similarity 0.83)
- Learning Visual Prompt for Gait Recognition (CVPR 2024 · similarity 0.83)
- Cross-Covariate Gait Recognition: A Benchmark (AAAI 2024 · similarity 0.82)
- Understanding Emotional Body Expressions via Large Language Models (AAAI 2025 · similarity 0.82)
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.