Research map / Architectures and efficiency / Attention and transformers
SDPose: Tokenized Pose Estimation via Circulation-Guide Self-Distillation
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 SDPose, a self-distillation method that loops tokens through transformer layers (Multi-Cycled Transformer) during training and distills the later cycles into a single-pass tiny transformer for human pose estimation.
Problem
Small transformer-based human pose estimators under-fit and perform worse than large models, while large ones are too costly for edge deployment; prior distillation needs a separate large teacher and extra feature alignment.
Approach
Increase effective latent depth without extra parameters by cycling tokens through the same transformer layers multiple times, and then self-distill the later-cycle tokens and predictions into the first cycle so that inference uses a single pass with no extra compute.
Key results
- SDPose-T reaches 69.7% AP on MSCOCO val with 4.4M params and 1.8 GFLOPs, matching TokenPose-S-V1 AP at 33% fewer params.
- SDPose-S-V2 reaches 73.5% AP on MSCOCO val with 6.2M params and 4.7 GFLOPs, versus 71.8% for TokenPose-S-V2.
- SDPose-S-V1 improves over TokenPose-S-V1 by 2.8% AP (72.3 vs 69.5) at the same params and GFLOPs.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace TokenPose-S-V2 with SDPose self-distillation | TokenPose-S-V2 · MSCOCO val2017 | AP: 71.8 -> 73.5 | Helped |
| replace TokenPose-S-V1 with SDPose self-distillation | TokenPose-S-V1 · MSCOCO val2017 | AP: 69.5 -> 72.3 | Helped |
| replace TokenPose-S-V1 with the smaller SDPose-T | TokenPose-S-V1 · MSCOCO val2017 | AP: 69.5 -> 69.7 | Helped |
| replace TokenPose-S-V2 with SDPose-S-V2 | TokenPose-S-V2 · Crowdpose test | AP: 62.3 -> 64.5 | Helped |
| add all three losses, Lpose, Lkt, and Lvt (ablation) | SDPose-S-V1 · MSCOCO val2017 | AP: 55.4% -> 72.3% | Helped |
| use 12 transformer layers for 2 cycles instead of 1 cycle, without distillation (ablation) | TokenPose-S-V1 · MSCOCO val2017 | AP: 69.5 -> 73.3(↑3.8%) | Helped |
| use 12 transformer layers for 3 cycles instead of 1 cycle, without distillation (ablation) | TokenPose-S-V1 · MSCOCO val2017 | AP: 69.5 -> 72.4(↑2.9%) | Helped |
| reduce the transformer layers from 12 to 4 and use 2 cycles (ablation) | TokenPose-S-V1 · MSCOCO val2017 | AP: 69.5 -> 69.4(↓0.1%) | Hurt |
| reduce the transformer layers from 12 to 4 and use 3 cycles (ablation) | TokenPose-S-V1 · MSCOCO val2017 | AP: 69.5 -> 70.8(↑1.3%) | Helped |
| use 3 cycles with distillation rather than 2 cycles (ablation) | SDPose-S-V1 · MSCOCO val2017 | AP: 72.3 -> 71.7 | Hurt |
| combine SDPose with PPT token pruning (ablation) | SDPose-S-V1 · MSCOCO val2017 | AP: 69.8 -> 72.3 | Helped |
| apply SDPose self-distillation to DeiT-Tiny (ablation) | DeiT-Tiny · ImageNet 1K | AP: 72.2 -> 72.7(↑0.5%) | Helped |
Surprising findings
- Using only the first cycle's output directly, without distillation, drastically reduces performance (55.4% AP in the ablation).
- Three cycles with distillation performed worse than two cycles, suggesting too many cycles cause loss of local keypoint information.
Limitations
- Too many cycles (three) degrade performance, indicating a limit on the cycling depth.
- Training cost is increased because the multi-cycle forward pass is used during training.
- Evaluation is mainly on MSCOCO and CrowdPose with the MMPose framework; generalization beyond these is limited to a single classification experiment.
Experimental setup
Datasets: MSCOCO train2017 CrowdPose ImageNet-1K
Benchmarks: MSCOCO val2017 MSCOCO test-dev CrowdPose test ImageNet-1K
Metrics: AP AR AP50 AP75 APM APL Params (M) GFLOPs FPS
Baselines: TokenPose-S-V1 TokenPose-S-V2 TokenPose-B OKDHP PPT-S PPT-B SimpleBaseline PRTR Poseur RLE DistilPose-S
Builds on: TokenPose DistilPose PPT Born-Again Neural Networks Be Your Own Teacher
Closest papers by meaning
- Hourglass Tokenizer for Efficient Transformer-Based 3D Human Pose Estimation (CVPR 2024 · similarity 0.85)
- Multiple View Geometry Transformers for 3D Human Pose Estimation (CVPR 2024 · similarity 0.85)
- SHaRPose: Sparse High-Resolution Representation for Human Pose Estimation (AAAI 2024 · similarity 0.85)
- Stable-Pose: Leveraging Transformers for Pose-Guided Text-to-Image Generation (NeurIPS 2024 · similarity 0.85)
- FoundationPose: Unified 6D Pose Estimation and Tracking of Novel Objects (CVPR 2024 · similarity 0.85)
- ExtPose: Robust and Coherent Pose Estimation by Extending ViTs (ICML 2025 · similarity 0.84)
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.