Research map / Video / Video segmentation and tracking
SEAS: ShapE-Aligned Supervision for Person Re-Identification
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 SEAS, which uses implicit 3-D body shape features (PIFu, plus SMPL for video) as pixel-level supervision during training to enrich ResNet-50/PSTA appearance features for frame- and video-based person re-identification, adding no inference cost.
Problem
Appearance-based person re-identification is affected by background variation because appearance lacks body shape awareness, and existing methods that add shape or gait encode the extra modality separately, adding parameters and cost.
Approach
Use body shape as training-time supervision rather than model input: a trainable deconvolutional extractor decodes the identity feature map into pixel-level PIFu features and is trained with a smooth L1 loss; for video, per-point feature variance across frames is minimized on a shared SMPL body model. The extractor is discarded at inference.
Key results
- On Market1501, ResNet-50 with SEAS reaches 98.6% rank-1 and 98.9% mAP, versus 96.9% prior best rank-1 (SOLIDER).
- On MARS, PSTA with SEAS reaches 95.1% rank-1 and 96.6% mAP, versus 91.5% for the PSTA baseline.
- On LS-VID, PSTA with SEAS and calibration reaches 90.5% rank-1 and 93.4% mAP, exceeding CAViT's 89.2% and 79.2%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply SEAS to ResNet-50 | ResNet-50 · Market1501 | Rank-1 accuracy: 94.1 -> 98.6 | Helped |
| Apply SEAS to ResNet-50 | ResNet-50 · MSMT17 | mAP: 47.2 -> 92.8 | Helped |
| Apply SEAS with feature calibration | PSTA · MARS | Rank-1 accuracy: 91.5 -> 95.1 | Helped |
| Apply SEAS with feature calibration | PSTA · LS-VID | mAP: 67.2 -> 93.4 | Helped |
| Add PIFu features as a second branch and concatenate with appearance (ablation) | ResNet-50 · Market1501 | Rank-1 accuracy: +0.0 | No clear effect |
| Add PIFu features as a second branch and concatenate with appearance (ablation) | ResNet-50 · Market1501 | mAP: +1.6 | Helped |
| Concatenate PIFu features channel-wise with appearance as input (ablation) | ResNet-50 · Market1501 | mAP: +2.6 | Helped |
| Use SPIN body-shape features as image-level supervision (ablation) | ResNet-50 with SEAS · Market1501 | mAP: +14.6 | Helped |
| Use PIFu pixel-level body-shape features as supervision instead of SPIN features (ablation) | ResNet-50 with SEAS · Market1501 | mAP: 97.8 -> 98.9 | Helped |
| Remove cross-frame feature calibration (ablation) | PSTA with SEAS · MARS | Rank-1 accuracy: 95.1 -> 94.8 | Hurt |
| Add SEAS (ablation) | BoT · Market1501 | Rank-1 accuracy: 94.5 -> 95.9 | Helped |
| Add SEAS (ablation) | LDS · MSMT17 | mAP: 67.2 -> 90.1 | Helped |
Surprising findings
- Using body shape as input (PIFu as a second branch or concatenation) adds parameters and FLOPs and gives smaller gains than using the same shape as supervision.
- Pixel-level PIFu supervision outperforms image-level SMPL (SPIN) supervision on Market1501 (98.6 vs 97.1 rank-1).
Limitations
- Performance depends on the quality of the pretrained shape encoder used for supervision (PIFu).
- PIFu reconstructions are limited by low image quality in re-identification datasets.
- Video-based training is expensive, taking one to three days depending on dataset.
Experimental setup
Datasets: Market1501 MSMT17 MARS LS-VID
Benchmarks: Market1501 MSMT17 MARS LS-VID
Metrics: rank-1 accuracy mAP
Baselines: SOLIDER PASS ViT-B TransReID SAN ASSP 3DInvarReID CAViT PSTA LDS STMN BiCnet-TKS BoT
Builds on: PIFu PSTA SPIN SMPL BoT
Closest papers by meaning
- Re-Identification Supervised Texture Generation (CVPR 2019 · similarity 0.87)
- Fine-Grained Shape-Appearance Mutual Learning for Cloth-Changing Person Re-Identification (CVPR 2021 · similarity 0.87)
- Distilling CLIP with Dual Guidance for Learning Discriminative Human Body Shape Representation (CVPR 2024 · similarity 0.86)
- Learning 3D Shape Feature for Texture-Insensitive Person Re-Identification (CVPR 2021 · similarity 0.85)
- Semantics-Aligned Representation Learning for Person Re-Identification (AAAI 2020 · similarity 0.85)
- From Poses to Identity: Training-Free Person Re-Identification via Feature Centralization (CVPR 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.