Source-linked AI summary
Deep Representation Learning with Part Loss for Person Re-Identification
Hantao Yao, Shiliang Zhang, Yongdong Zhang, Jintao Li, Qi Tian
TL;DR
Person ReID needs representations that distinguish unseen identities, but classification-trained networks may focus on only training-discriminative body regions. PL-Net jointly minimizes global classification loss and an automatically generated multi-part loss, and experiments on three datasets show promising performance against existing deep representations.
Problem
Person ReID is a zero-shot problem, but classification-trained representations optimize seen-person classification and may lack descriptions of other parts useful for unseen identities.
Method
PL-Net minimizes global classification loss together with part loss, which generates K body parts unsupervised and applies person classification loss separately to each part.
Results
PL-Net learns more reliable representations and achieves promising performance on Market1501, CUHK03, and VIPeR; on Market1501 it reaches 69.3% mAP and 88.2% Rank-1 accuracy.
Takeaways & Limitations
Considering multiple body-part losses encourages discriminative representations for different parts rather than relying only on globally classification-discriminative regions.
Takeaways & Limitations
Person ReID assumes that probe identities are absent from the training set and gallery identities are distinct from training identities.
Abstract
from arXiv · showhide
Learning discriminative representations for unseen person images is critical for person Re-Identification (ReID). Most of current approaches learn deep representations in classification tasks, which essentially minimize the empirical classification risk on the training set. As shown in our experiments, such representations commonly focus on several body parts discriminative to the training set, rather than the entire human body. Inspired by the structural risk minimization principle in SVM, we revise the traditional deep representation learning procedure to minimize both the empirical classification risk and the representation learning risk. The representation learning risk is evaluated by the proposed part loss, which automatically generates several parts for an image, and computes the person classification loss on each part separately. Compared with traditional global classification loss, simultaneously considering multiple part loss enforces the deep network to focus on the entire human body and learn discriminative representations for different parts. Experimental results on three datasets, i.e., Market1501, CUHK03, VIPeR, show that our representation outperforms the existing deep representations.
1. Introduction
Person ReID is a zero-shot identification problem requiring representations that distinguish unseen people. The paper argues that classification-trained networks overfocus on training-discriminative body regions and proposes PL-Net to learn more comprehensive part-aware representations.
- Person ReID requires discriminative representations because training and test sets contain no person identities in common.
- Classification-trained representations optimize training-set classification risk, which may not produce representations optimal for distinguishing unseen person images.
- Classification networks mainly focus on one discriminative body region, such as the upper body, while ignoring other potentially meaningful parts.
- PL-Net evaluates representation learning risk with part loss by generating K image parts and computing person classification loss separately for each part.
- PL-Net jointly minimizes global classification loss and part loss, producing reliable representations and promising performance across Market1501, CUHK03, and VIPeR.
- The paper frames representation learning risk and part loss as an original response to the limitations of directly training deep classification models for person ReID.
2. Related Work
Prior person ReID research uses classification, siamese, triplet, and local-part approaches to learn representations or similarities. PL-Net differs from local-part methods by automatically detecting human parts without extra annotations or detectors.
- Deep classification models have been widely fine-tuned or trained to extract person ReID representations.
- Siamese networks learn from image pairs to verify similarity, while related methods infer descriptions or similarity metrics jointly.
- Recent local-part methods use body-joint annotations, detectors, or spatial transformation for part localization and feature fusion.
- PL-Net automatically detects human parts without extra annotations or detectors, making it more efficient and easier to implement than these local-part algorithms.
3. Methodology
PL-Net addresses the mismatch between training-set classification and person ReID’s unseen identities by learning representations from both global and body-part classification losses. It automatically generates parts from feature-map activations and jointly optimizes their losses with a global loss.
- Person ReID is treated as zero-shot recognition because training and gallery sets contain distinct identities, requiring representations for unseen people.
- Training classification models may focus on discriminative regions such as the upper body while neglecting head, lower-body, and foot cues useful for unseen identities.
- PL-Net evaluates representation-learning risk by computing person classification loss separately on K automatically generated body parts.
- The baseline replaces fully connected layers with a convolutional layer and Global Average Pooling, whose parameter-free classifier encourages stronger feature learning and reduces overfitting.
- PL-Net jointly minimizes global classification loss and part loss, while updating part generation during training; with K=4, boxes coarsely cover major body regions, and K=8 captures more detailed parts.
- Feature maps are clustered by the vertical locations of their maximum activations, then converted into part boxes through pooled saliency maps, thresholding, and enclosing rectangles.
4. Experiments
Experiments evaluate PL-Net across three person ReID datasets and test generated parts, part loss, representation composition, and parameter K. The results show improved part and global representations, with strong comparisons against existing methods.
- Datasets: PL-Net is evaluated on VIPeR, CUHK03, and Market1501 using dataset-specific cross-camera and retrieval metrics.VIPeR and CUHK03 use Rank-1, Rank-5, and Rank-10 accuracy; Market1501 uses Rank-1 accuracy and mAP.
- Part generation: Generated parts achieve substantially higher accuracy than fixed grid parts for K = 4 and K = 8 on Market1501.The authors attribute this to better human-body coverage and filtering of clustered backgrounds.
- Part loss: Part loss improves the discriminative performance of part representations with both fixed grid parts and generated parts.The fixed-grid experiment uses K = 4 and 8, while the generated-part experiment reports consistent improvements.
- Global representation: Part loss also boosts the global representation on Market1501, with mAP and Rank-1 accuracy increasing as K changes from the no-part-loss setting.K = 0 denotes the setting without generated parts or part loss.
- Final representation: The final representation performs better with larger K, leading the authors to set K = 8 for subsequent experiments.Larger K extracts more detailed parts, consistent with the global-representation results.
- Part-loss design: Computing classification error separately on each part outperforms computing it after concatenating part features, achieving 67.17% versus 64.72%.The separate computation is reported as more effective for ensuring the quality of each learned part feature.
- State-of-the-art comparison: On Market1501, PL-Net reaches 69.3% mAP and 88.2% Rank-1 accuracy, while part loss improves baseline global and part mAP by 4% and 7.1%.Combining global and part representations further boosts performance.
- State-of-the-art comparison: On CUHK03, PL-Net achieves 82.75%, 96.59%, and 98.59% Rank-1, Rank-5, and Rank-10 accuracy, respectively.The global and part representations improve the baseline Rank-1 accuracy by 8.1% and 9.85%, respectively.
5. Conclusions
The paper proposes part loss to reduce representation learning risk for unseen person images by learning discriminative representations across automatically generated body parts.
- Part loss evaluates representation learning risk by generating K body parts unsupervised and optimizing classification loss separately for each part.This extends training beyond a single global classification objective.
- The method targets person ReID as a zero-shot setting, where representations must remain discriminative for unseen person images.
- Optimizing separate part losses helps the network learn discriminative representations for different body parts.
- Experiments on Market1501, CUHK03, and VIPeR demonstrate advantages over existing deep representations.
- The approach explicitly infers parts using a specified parameter K, while more implicit approaches are left for future work.