Research map / Label-efficient and robust learning / Recognition and long-tail
Pedestrian Attribute Recognition as Label-balanced Multi-label Learning
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.
placeholder
Problem
Pedestrian attribute recognition (PAR) datasets are heavily label-imbalanced and co-occurrence entangled, so image-level label-balanced re-sampling is infeasible, and feature augmentation for under-represented attributes overfits.
Approach
Decouple feature extractor (instance-balanced training) from attribute-wise classifiers retrained on label-balanced re-sampled cached features (FRDL); augment features by translating them along their local loss gradient with SGD trajectories, giving heterogeneous Bayesian-like sampling (GOAT), plus a feature de-noising loss.
Key results
- FRDL+GOAT reaches 89.44% mA on PA100k and 87.72% mA on RAPv1, the best among compared methods in the paper's table.
- On UPAR*, the method reports 79.40 mA versus 72.58 for PARFormer-B, and on PETAzs 79.10 mA versus 76.16.
- Re-weighting alone gives about 1.5-3% mA gain over the baseline; FRDL alone reaches 88.53% mA on PA100k versus 82.30 baseline.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use label-balanced feature re-sampling and decoupled classifier training instead of the baseline | PAR · PA100k, RAPv1, PETA | mA: 82.30 -> 88.53; 80.61 -> 86.31; 86.29 -> 87.55 | Helped |
| Add GOAT feature augmentation to FRDL (ablation) | FRDL · PA100k, RAPv1, PETA | mA: 88.53 -> 89.44; 86.31 -> 87.72; 87.55 -> 88.59 | Helped |
| Replace ISDA covariance with random noise covariance (ablation) | FRDL + ISDA · PA100k, RAPv1, PETA | mA: 88.70 -> 88.79; 86.52 -> 86.46; 87.94 -> 87.93 | No clear effect |
| Add ISDA feature augmentation (ablation) | FRDL · PA100k, RAPv1, PETA | mA: 88.53 -> 88.70; 86.31 -> 86.52; 87.55 -> 87.94 | Helped |
| Add MetaSAug feature augmentation (ablation) | FRDL · PA100k, RAPv1, PETA | mA: 88.53 -> 88.63; 86.31 -> 86.58; 87.55 -> 87.80 | Helped |
| Generate GOAT features along an SGD trajectory rather than using Eq. [ref] translation alone (ablation) | FRDL + GOAT · PA100k, RAPv1, PETA | mA: 88.84 -> 89.15; 86.75 -> 87.52; 87.96 -> 88.27 | Helped |
| Add feature de-noising to GOAT features generated along an SGD trajectory (ablation) | FRDL + GOAT (SGD) · PA100k, RAPv1, PETA | mA: 89.15 -> 89.44; 87.52 -> 87.72; 88.27 -> 88.59 | Helped |
| Use GOAT rather than standard pixel-space augmentation (ablation) | PAR · PA100k, PETAzs, RAPzs | mA: 82.55 -> 84.16; 75.22 -> 76.41; 76.43 -> 77.98 | Helped |
| Use Cutout augmentation (ablation) | PAR baseline · PA100k, PETAzs, RAPzs | mA: 82.45 -> 81.09; 75.01 -> 73.83; 76.14 -> 73.56 | Hurt |
| Use Mixup augmentation (ablation) | PAR baseline · PA100k, PETAzs, RAPzs | mA: 82.45 -> 82.07; 75.01 -> 74.50; 76.14 -> 75.79 | Mixed |
| Add dropout variational inference to GOAT (ablation) | GOAT · PA100k, PETAzs, RAPzs | mA: 84.16 -> 84.50; 76.41 -> 76.66; 77.98 -> 78.41 | Helped |
Surprising findings
- Homogeneous feature augmentation (ISDA) with random covariance performs about as well as the original ISDA, suggesting the covariance design matters little once margin is tuned.
- F1 and mA are inconsistent: the method does not top F1 on several benchmarks even though it leads on mA.
Limitations
- Attribute F1 lags behind some methods despite leading mA.
- Hyperparameters such as the SGD trajectory length T and step size were chosen per setting; large T risks out-of-distribution features.
Experimental setup
Datasets: PA100k PETA RAPv1 PETAzs RAPzs UPAR
Benchmarks: PA100k RAPv1 PETA PETAzs RAPzs UPAR*
Metrics: mA F1
Baselines: ResNet-50 baseline ConvNeXt-base baseline DAFL DRFormer IAA-Caps EALC DFDT FEMDAR VAC JLAC L2L OAGCN
Builds on: decoupled learning ISDA MetaSAug class-balanced loss re-weighting
Closest papers by meaning
- Pedestrian Attribute Recognition: A New Benchmark Dataset and a Large Language Model Augmented Framework (AAAI 2025 · similarity 0.88)
- Selective and Orthogonal Feature Activation for Pedestrian Attribute Recognition (AAAI 2024 · similarity 0.86)
- Multispectral Pedestrian Detection with Sparsely Annotated Label (AAAI 2025 · similarity 0.83)
- Enhanced Visual-Semantic Interaction with Tailored Prompts for Pedestrian Attribute Recognition (CVPR 2025 · similarity 0.83)
- Attribute-Guided Pedestrian Retrieval: Bridging Person Re-ID with Internal Attribute Variability (CVPR 2024 · similarity 0.83)
- Optimizing Label Assignment for Weakly Supervised Person Search (AAAI 2025 · 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.