Research map / Label-efficient and robust learning / Recognition and long-tail
Selective and Orthogonal Feature Activation for Pedestrian Attribute 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 SOFAFormer, a ViT-based pedestrian attribute recognition model with a parameter-free Selective Feature Activation Method (SFAM) and an Orthogonal Feature Activation Loss (LOFA) to counter attribute imbalance.
Problem
Pedestrian attribute recognition struggles with CNNs missing long-range inter-attribute relations and with attribute imbalance, where frequent attributes dominate and rarer attributes (e.g., glasses, backpack) are overlooked.
Approach
Use a ViT backbone and, without extra parameters, suppress the most activated spatial regions of the final feature map (SFAM) so the classifier must learn from under-activated attribute regions; a loss (LOFA) pushes the original and suppressed features toward orthogonality to keep them complementary.
Key results
- On PA100K, SFAM raises mA from 81.4% to 82.9% and adding LOFA reaches 83.4% mA, the best reported result.
- SOFAFormer outperforms the second-best method by 0.4% mA on PETA, 1.2% on PA100K, 1.3% on RAPv1, and 1.9% on RAPv2.
- On the zero-shot-style PETAzs and RAPv2zs splits with non-overlapping pedestrian IDs, it beats the best competitor IAA-Caps by 2.2% and 1.9% mA.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use SOFAFormer instead of GRL | Pedestrian attribute recognition · PETA | mA: 86.7 -> 87.1 | Helped |
| use SOFAFormer instead of Label2Label | Pedestrian attribute recognition · PA100K | mA: 82.2 -> 83.4 | Helped |
| use SOFAFormer instead of SSChard | Pedestrian attribute recognition · RAPv1 | mA: 82.1 -> 83.4 | Helped |
| use SOFAFormer instead of IAA-Caps | Pedestrian attribute recognition · RAPv2 | mA: 80.0 -> 81.9 | Helped |
| add SFAM (ablation) | SOFAFormer · PA100K | mA: 81.4 -> 82.9 | Helped |
| add LOFA (ablation) | SOFAFormer with SFAM · PA100K | mA: 82.9 -> 83.4 | Helped |
| add SFAM (ablation) | SOFAFormer · RAPv2 | mA: 79.6 -> 81.0 | Helped |
| add LOFA (ablation) | SOFAFormer with SFAM · RAPv2 | mA: 81.0 -> 81.9 | Helped |
Surprising findings
- Setting the suppression ratio alpha to 0 (fully erasing the top activations) makes training loss become NaN.
Limitations
- Hyperparameters alpha and tau were chosen on PA100K and the mFive criterion, with sensitivity shown only on that dataset.
- Relies on a ViT backbone, so its gains are not isolated from backbone size effects in the comparison.
Experimental setup
Datasets: PETA PA100K RAPv1 RAPv2
Benchmarks: PETA PA100K RAPv1 RAPv2 PETAzs RAPv2zs
Metrics: mA accuracy precision recall F1
Baselines: GRL IAA-Caps ALM SSC Label2Label JLAC VRKD VAC
Builds on: Vision Transformer Zagoruyko attention transfer activation mapping Grad-CAM
Closest papers by meaning
- Pedestrian Attribute Recognition: A New Benchmark Dataset and a Large Language Model Augmented Framework (AAAI 2025 · similarity 0.93)
- Learning Disentangled Attribute Representations for Robust Pedestrian Attribute Recognition (AAAI 2022 · similarity 0.89)
- Enhanced Visual-Semantic Interaction with Tailored Prompts for Pedestrian Attribute Recognition (CVPR 2025 · similarity 0.88)
- RGB-Event based Pedestrian Attribute Recognition: A Benchmark Dataset and An Asymmetric RWKV Fusion Framework (CVPR 2026 · similarity 0.87)
- Pedestrian Attribute Recognition as Label-balanced Multi-label Learning (ICML 2024 · similarity 0.86)
- Attribute-Guided Pedestrian Retrieval: Bridging Person Re-ID with Internal Attribute Variability (CVPR 2024 · similarity 0.86)
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.