Localization Guided Learning for Pedestrian Attribute Recognition

1 Introduction

• Pedestrian attribute recognition is used in person re-identification, retrieval, and social behavioral analysis, and deep learning methods have advanced beyond earlier SVM-based approaches.Recent methods frame recognition as multi-label classification and use parameter-sharing networks to exploit semantic relations among attributes.Source: §1 · p. 1
• Small attributes such as glasses and shoes can be difficult to classify after down-sampling, while pedestrian images vary in shape and viewpoint.Backpacks may also appear at different relative positions, making localization by fixed relative position difficult.Source: §1 · pp. 1–2
• LG-Net uses class activation maps to localize attributes, weights local features attribute-wise, and merges them with global features for prediction.The authors report evaluation on two large pedestrian datasets using five metrics, with significant improvement in attribute prediction.Source: §1 · p. 3
• Part- and pose-based methods use human-defined local feature strategies, while LG-Net targets attribute-specific locations and surrounding local features.The authors describe LG-Net as achieving state-of-the-art results on the two largest pedestrian attribute datasets.Source: §1 · pp. 2–3

Pedestrian Attribute Recognition

• Deep learning methods moved pedestrian attribute recognition beyond traditional single-attribute SVMs toward shared networks that exploit semantic relations among attributes.Sudowe et al. use independent loss layers for each attribute in an end-to-end network.Source: §2 · p. 3
• Other approaches capture multi-level features, address sample imbalance, use viewpoint guidance, or model intra-person, inter-person, and inter-attribute context.These approaches are attributed to Yu et al. [32], Liu et al. [20], Li et al., Sarfraz et al. [25], and Wang et al..Source: §2 · p. 3
• Pose and part methods use poselets, fixed image splits, or keypoints to extract local features, but the cited approaches lack attribute-specific localization.The authors connect this missing localization information with lower robustness to changes in objects’ relative positions.Source: §2 · pp. 3–4

Weakly Supervised Localization

• Weakly supervised localization seeks object localization using only image-level supervision, with approaches including multiple-instance learning, CAMs, region selection, and spatial hiding.The section describes bounding boxes over CAM highlight areas as one localization approach.Source: §2 · p. 4

3 Localization Guided Network Architecture

• LG-Net has separate global and local branches, adapted from Inception-v2, for generating attribute locations and using them in attribute prediction.The global branch takes the whole image and produces class activation maps for attribute localization.Source: §3 · p. 4, §3.1 · p. 4

3.1 Global Feature Extraction

• A fixed 1×1 convolution initialized from a pretrained classifier generates one class activation map per attribute, whose high-response regions form activation boxes.The global branch parameters remain fixed so the activation maps and resulting boxes serve as localization guidance.Source: §3.1 · p. 4, Figure 2 · p. 5

3.2 Local Feature Extraction

• EdgeBoxes supplies candidate regions, and ROI pooling extracts proposal features between Inception-5a and Inception-5b for later use.A fixed number of proposals is selected; the proposal features are processed into separate feature vectors.Source: §3.2 · p. 5

3.3 Localization Guidance Module

• The localization guidance module weights proposal features by their IoU with attribute-specific CAM boxes, then combines attribute-wise local features with global features.The affinity map is linearly normalized; local features are aggregated into attribute-wise vectors and fused by element-wise summation.Source: §3.3 · pp. 6–7

3.4 Implementation Details

• Training proceeds in two stages: first train an Inception-v2 classifier, then initialize and freeze the global branch from its fully connected weights.The method selects the top 100 EdgeBoxes proposals per image and defines activated regions using a threshold of 0.2 times the image’s maximum activation for that attribute.Source: §3.4 · p. 7

4.1 Benchmark Data

• RAP contains 41,585 indoor-surveillance images, with 51 of its 72 attributes selected for evaluation; PA-100K contains 100,000 outdoor-surveillance images labeled with 26 binary attributes.Both datasets are blurry due to relatively low resolution, and each binary attribute has a low positive ratio; evaluation uses mA and sample-based accuracy, precision, recall, and F1.Source: §4.1 · pp. 7–8

4.2 Quantitative Comparison with Prior Arts

• LG-Net surpasses previous state-of-the-art methods on both RAP and PA-100K across all five metrics.The comparison follows prior evaluation settings and includes an Inception-v2 baseline.Source: Table 1 · p. 8, §4.2 · p. 8

4.3 Qualitative Evaluation

• On RAP, LG-Net performs better than VeSPA on most of 51 attributes, especially attached-object attributes such as backpacks, plastic bags, and calling.The authors also report slightly better action recognition for talking and gathering, attributing this to the use of both global and local information.Source: §4.3 · p. 8, Figure 3 · p. 9

Mapping Resolution

• The 14×14 feature-map model surpasses the 7×7 baseline by 0.92% in mA and 0.45% in Accuracy.The higher-resolution maps incur minor computational inefficiency and apparently improve localization.Source: §5 · p. 9, Table 2 · p. 10

Dilated Convolution

• The final model adopts dilated convolutions with dilation ratios 1 and 2.The study compares ordinary convolution with dilation ratios 1 and 2, including a variant that applies ratio 2 on one branch only.Source: §5 · p. 9

Affinity Strategy

• IoU proposal weighting boosts mA by 0.84% and Accuracy by 0.42% over weighting by overlapping area.The adopted strategy computes affinity using Intersection over Union.Source: §5 · p. 9, Table 2 · p. 10

Localization Guidance

• Removing localization guidance reduces accuracy by 4.4% compared with the original LG-Net.The modified model removes the localization generator and guidance module and weights all proposals evenly.Source: §5 · p. 10, Table 2 · p. 10

• The network achieves state-of-the-art results on the two largest pedestrian attribute datasets.Source: §6 · p. 10
• Ablation studies identify localization guidance as the key factor in the localization network’s success.They also show the effectiveness of each network component.Source: §6 · p. 10
• LG-Net outputs localization results in parallel with attribute predictions.Source: §6 · p. 10
• The network uses attribute-specific localization guidance for local feature extraction and utilization.Its findings indicate that attribute spatial information is helpful to local-feature utility.Source: §6 · p. 10