Source-linked AI summary
Camera Style Adaptation for Person Re-identification
Zhun Zhong, Liang Zheng, Zhedong Zheng, Shaozi Li, Yi Yang
TL;DR
Person re-identification must handle appearance changes caused by different cameras. CamStyle uses CycleGAN to generate camera-style-transferred training images and applies LSR to reduce their noise, yielding consistent improvement across systems and up to 17.1% improvement in few-camera systems.
Problem
Person re-identification suffers from image style variations caused by different cameras in cross-camera retrieval.
Method
CamStyle uses CycleGAN to translate labeled training images into other camera styles, combines them with original images, and applies LSR to generated samples.
Results
CamStyle improves few-camera systems by as much as 17.1%, while LSR produces consistent improvement across systems.
Takeaways & Limitations
Camera-aware style adaptation can reduce over-fitting and support camera-invariant person descriptors without additional human labeling.
Takeaways & Limitations
Style-transferred images introduce noise from imperfect CycleGAN transfer and noisy real samples, compromising vanilla CamStyle when more camera data reduces over-fitting risk.
Abstract
from arXiv · showhide
Being a cross-camera retrieval task, person re-identification suffers from image style variations caused by different cameras. The art implicitly addresses this problem by learning a camera-invariant descriptor subspace. In this paper, we explicitly consider this challenge by introducing camera style (CamStyle) adaptation. CamStyle can serve as a data augmentation approach that smooths the camera style disparities. Specifically, with CycleGAN, labeled training images can be style-transferred to each camera, and, along with the original training samples, form the augmented training set. This method, while increasing data diversity against over-fitting, also incurs a considerable level of noise. In the effort to alleviate the impact of noise, the label smooth regularization (LSR) is adopted. The vanilla version of our method (without LSR) performs reasonably well on few-camera systems in which over-fitting often occurs. With LSR, we demonstrate consistent improvement in all systems regardless of the extent of over-fitting. We also report competitive accuracy compared with the state of the art.
1. Introduction
CamStyle explicitly adapts person re-ID training data to camera styles, using CycleGAN-generated images and LSR to address transfer noise. The method targets camera variation, data scarcity, and over-fitting, with reported gains in few-camera and full-camera systems.
- Person re-ID retrieves the same person across camera databases, where differing resolution and illumination produce substantial appearance and background variation.
- CamStyle augments labeled training data by translating images into other camera styles with CycleGAN while preserving their original identity labels.
- Style transfer reduces over-fitting and supports camera-invariant descriptors, but generated samples introduce noise, especially when more camera data lowers over-fitting risk.
- The vanilla camera-aware style-transfer model improves few-camera systems by as much as 17.1%.
- LSR applied to style-transferred samples yields consistent improvement in full-camera systems.
2. Related Work
Prior re-ID work develops deep representations and classification-based descriptors, while related GAN research provides unpaired image-to-image translation and style-transfer foundations. CamStyle builds on these directions for camera-aware training-data generation.
- Deep learning person re-identification: Deep re-ID methods use siamese CNNs, deeper networks, or LSTM-enhanced architectures to learn robust and discriminative features.
- Deep learning person re-identification: Classification-based approaches exploit re-ID labels, including IDE, Feature Fusion Net, and SVDNet with orthogonal fully connected weights.
- Deep learning person re-identification: Data augmentation and regularization methods generate varied samples or erase image regions to improve generalization and robustness to occlusion.
- Generative Adversarial Networks: GAN research includes image-to-image translation, style transfer, and cross-domain generation, with CycleGAN enabling translation between unpaired domains.
- Generative Adversarial Networks: CamStyle applies camera-aware style transfer to re-ID training by combining generated images with real images in its CNN-training pipeline.
3. The Proposed Method
CamStyle uses CycleGAN to transfer training images between camera styles, augmenting re-ID data while preserving identity labels. Because translation can introduce noise, the method uses label smoothing regularization to reduce its impact during training.
- Camera-aware image translation: CycleGAN learns image-to-image translation models between camera domains, generating style-transferred training images from original samples.The method also adds an identity mapping loss to encourage color consistency between input and output images.
- Data augmentation mechanism: Style-transferred samples generally overlap real samples in feature space, but occasional noisy samples appear and require regularization.The visualization uses dots for real samples, triangles for fake samples, and colors for identities.
- Camera-aware image translation: Each training image produces L −1 samples styled like the other cameras, and generated images retain the original identity labels.The augmented set combines original and style-transferred images for re-ID CNN training.
- Training with CamStyle: The vanilla version trains real and fake samples with one-hot identity labels, whereas the full version applies label smooth regularization to account for fake-sample noise.For real images, labels are treated as correct; for style-transferred images, label distributions are softened.
- Training with CamStyle: The vanilla approach improves baseline IDE accuracy in few-camera systems, where over-fitting tends to occur, but its effectiveness is compromised as camera data become more abundant.With more cameras, over-fitting is less critical and transfer noise becomes more prominent.
- Training with CamStyle: Transfer noise comes from imperfect CycleGAN generation and from propagating noisy real samples caused by occlusion or detection errors.Such noisy generated samples reduce the benefit of augmentation in full-camera systems with lower over-fitting risk.
4. Experiment
Experiments on Market-1501 and DukeMTMC-reID evaluate CamStyle across camera counts, fake-to-real sampling ratios, loss functions, augmentation combinations, and state-of-the-art comparisons. CamStyle is especially effective in few-camera systems, while LSR improves performance consistently and yields competitive full-camera results.
- Parameter Analysis: CamStyle consistently improves over the baseline across different ratios of real and fake samples, with best performance at M : N = 3 : 1.Using more fake than real data yields about 1% rank-1 improvement, whereas M : N > 1 yields more than 2%.
- Variant Evaluation: 17.1% rank-1 improvement is achieved by vanilla CamStyle on Market-1501’s 2-camera system, increasing accuracy from 43.2% to 60.3%.The corresponding DukeMTMC-reID 2-camera result improves from 45.3% to 54.8%.
- Variant Evaluation: 88.12% rank-1 accuracy is reached on the full-camera Market-1501 system when LSR replaces cross-entropy on style-transferred data.Cross-entropy on the fake data reaches 86.31%, while LSR on the fake data reaches 88.12%; LSR alone on real data does not help much or can reduce performance.
- Variant Evaluation: 89.49% and 78.32% rank-1 accuracy are obtained on Market-1501 and DukeMTMC-reID, respectively, after combining CamStyle with Random Erasing.CamStyle is complementary to the other evaluated augmentation methods, and the three-method combination performs best in the reported experiment.
- Comparison with State-of-the-art Methods: CamStyle achieves 88.12% rank-1 accuracy on Market-1501 and 75.27% on DukeMTMC-reID, producing competitive results against state-of-the-art methods.The paper also reports lower mAP than TriNet by 0.42% on Market-1501 and than SVDNet by 3.32% on DukeMTMC-reID.
5. Conclusion
CamStyle uses CycleGAN to generate camera-style-transferred training images, while LSR reduces the noise those samples introduce. On Market-1501 and DukeMTMC-reID, the combined method consistently improves over baselines and complements other augmentation techniques.
- CamStyle learns camera-pair style-transfer models with CycleGAN to generate new training images from originals.The original and style-transferred images form the new training set.
- Label smooth regularization alleviates noise induced by CycleGAN-generated samples.
- Experiments on Market-1501 and DukeMTMC-reID show that CamStyle reduces the impact of over-fitting.
- Combining CamStyle with LSR yields consistent improvement over the baselines.
- CamStyle is complementary to other data augmentation techniques.