Source-linked AI summary
Beyond triplet loss: a deep quadruplet network for person re-identification
Weihua Chen, Xiaotang Chen, Jianguo Zhang, Kaiqi Huang
TL;DR
Person ReID must generalize from training identities to unseen test identities, but triplet loss primarily enforces training-set ranking and can leave substantial intra-class variation. The paper introduces a quadruplet loss and a margin-based online hard-negative-mining network; it reports stronger variation structure and performance than most state-of-the-art methods on representative datasets.
Problem
Triplet loss focuses on correct training-set ordering, while ReID testing identities are unseen and the resulting model can generalize weakly.
Method
The paper combines a quadruplet ranking loss with a quadruplet deep network using margin-based online hard negative mining.
Results
The proposed network outperformed most state-of-the-art methods on CUHK03, CUHK01, and VIPeR, while producing larger inter-class and smaller intra-class variations.
Takeaways & Limitations
The quadruplet design improves the reported testing-set performance by targeting inter-class and intra-class variation beyond triplet ranking.
Takeaways & Limitations
The adaptive margin assumes that distances or similarities between same-class positive pairs follow the distributional behavior used for sampling.
Abstract
from arXiv · showhide
Person re-identification (ReID) is an important task in wide area video surveillance which focuses on identifying people across different cameras. Recently, deep learning networks with a triplet loss become a common framework for person ReID. However, the triplet loss pays main attentions on obtaining correct orders on the training set. It still suffers from a weaker generalization capability from the training set to the testing set, thus resulting in inferior performance. In this paper, we design a quadruplet loss, which can lead to the model output with a larger inter-class variation and a smaller intra-class variation compared to the triplet loss. As a result, our model has a better generalization ability and can achieve a higher performance on the testing set. In particular, a quadruplet deep network using a margin-based online hard negative mining is proposed based on the quadruplet loss for the person ReID. In extensive experiments, the proposed network outperforms most of the state-of-the-art algorithms on representative datasets which clearly demonstrates the effectiveness of our proposed method.
1. Introduction
Person ReID must generalize across unseen identities despite large appearance variation. The paper addresses triplet-loss generalization by introducing a quadruplet loss and network designed to reduce intra-class variation and enlarge inter-class variation.
- Person ReID identifies people across cameras despite appearance changes caused by poses, illumination, and viewpoints.
- Triplet loss mainly enforces correct identity ranking on training probes, but testing identities are unseen and generalization can be weak.
- The proposed quadruplet loss further reduces intra-class variation and enlarges inter-class variation while preserving correct probe-image rankings.
- The paper reports theoretical loss analysis and superior performance to most state-of-the-art methods on CUHK03, CUHK01, and VIPeR.
- The quadruplet network uses margin-based online hard negative mining, which adaptively sets a margin threshold to select hard training samples.
2. Related work
Earlier ReID methods separately address feature extraction or similarity measurement, while deep methods learn both through ranking or classification losses. The paper positions its quadruplet loss as retaining relative ranking relationships while additionally reducing intra-class variation.
- Traditional ReID systems often separate handcrafted feature extraction from similarity measurement, limiting end-to-end global optimization.
- Deep ReID approaches learn features and metrics end to end and commonly formulate the task through ranking or classification losses.
- DeepLDA and ImpTrpLoss also reduce intra-class variation, but DeepLDA partly ignores relative relationships between image pairs.
- The quadruplet loss expands the triplet loss, preserving relative relationships while adding a constraint involving pairs with different probe images.
- Existing sampling schemes use offline or predefined selection, whereas the proposed method adaptively selects hard samples using a model-dependent margin threshold.
3. The proposed approach
The proposed approach extends triplet-based ReID with a quadruplet loss and an adaptive margin-based online hard negative mining strategy. It adds constraints across different probe images while balancing stronger and weaker pushes to improve inter-class separation and sample selection.
- 3.1. The triplet loss: The triplet loss keeps positive pairs closer than negative pairs, using normalized features or a learned distance with a margin threshold.The learned metric can model complex cross-camera relationships, but its unbounded scalar output can partly invalidate the margin threshold.
- 3.2. The quadruplet loss: The quadruplet loss adds a second constraint that compares positive and negative pairs with different probe images.Its first term preserves same-probe ordering, while the second requires minimum inter-class distance to exceed maximum intra-class distance.
- 3.2. The quadruplet loss: The two quadruplet terms use different margin thresholds: a strong push for same-probe pairs and a weaker push for different-probe pairs.This design keeps the auxiliary cross-probe constraint from leading training while further enlarging inter-class variation.
- 3.3. Margin-based online hard negative mining: The adaptive margin is based on the mean distances of positive and negative pair distributions, with separate correlation settings for the two loss margins.The implementation estimates these means per batch and backpropagates samples whose distances are smaller than the current average.
4. Relationships of different losses
The paper analyzes triplet, binary classification, contrastive, and quadruplet losses in a unified framework for person ReID. It argues that quadruplet training preserves ranking relationships while improving separation between identities.
- The analysis compares the proposed quadruplet loss with triplet and binary classification losses under a unified theoretical framework.
- The paper presents binary classification's locally optimal wrong-order solution as undesirable for person ReID, where ranking relationships for the same probe matter most.
- The contrastive loss is transformed from doublets into quadruplets, splitting training into terms for shared-probe and different-probe pairs.
- Binary classification and contrastive losses can favor low misclassification or small positive-pair distances even when relative ranking errors remain.
- The quadruplet loss combines correct ordering for pairs sharing a probe with pushing negative pairs away from positive pairs across different probes.
5. Experiment
Experiments evaluate different losses, architectures, and state-of-the-art comparisons on CUHK03, CUHK01, and VIPeR. The proposed models improve distance distributions and outperform most compared ReID methods, with dataset-dependent performance differences.
- 5. Experiment: The experiments evaluate different losses and compare the proposed approach with state-of-the-art methods on three ReID datasets.
- 5.2. Results of Quadruplet Network: Table 1 compares triplet, classification, improved-triplet, quadruplet, and quadruplet-plus-hard-negative-mining architectures using CMC performance.
- 5.2. Results of Quadruplet Network: The quadruplet variants progressively reduce average intra-class distances and increase average inter-class distances on the CUHK03 training set.
- 5.3. Comparison with the state of the arts: The proposed multi-task network outperforms most existing methods in rank-1 accuracy across CUHK03, CUHK01, and VIPeR.
- 5.3. Comparison with the state of the arts: DGD achieves better rank-1 performance, while the proposed method achieves higher rank-n performance on CUHK03 for n>1.
- 5.3. Comparison with the state of the arts: VIPeR's relatively small size may limit the demonstrated potential of deep learning, making handcrafted metric learning more advantageous there.
6. Conclusion
The paper concludes that its quadruplet loss and margin-based online hard negative mining network outperform most state-of-the-art methods on three representative person ReID datasets.
- The quadruplet loss and network outperform most state-of-the-art methods on CUHK03, CUHK01, and VIPeR.