Sparse Label Smoothing Regularization for Person Re-Identification

Jean-Paul Ainam, Ke Qin, Guisong Liu, Guangchun LuoIEEE Access 2019 · OpenReview

Abstract: Person re-identification (re-id) is a cross-camera retrieval task which establishes a correspondence between images of a person from multiple cameras. Deep learning methods have been successfully applied to this problem and have achieved impressive results. However, these methods require a large amount of labeled training data. Currently, the labeled datasets in person re-id are limited in their scale and manual acquisition of such large-scale datasets from surveillance cameras is a tedious and labor-intensive task. In this paper, we propose a framework that performs intelligent data augmentation and assigns the partial smoothing label to generated data. Our approach first exploits the clustering property of existing person re-id datasets to create groups of similar objects that model cross-view variations. Each group is then used to generate realistic images through adversarial training. Our aim is to emphasize the feature similarity between generated samples and the original samples. Finally, we assign a non-uniform label distribution to the generated samples and define a regularized loss function for training. The proposed approach tackles two problems 1) how to efficiently use the generated data and 2) how to address the over-smoothness problem found in current regularization methods. The extensive experiments on four large-scale datasets show that our regularization method significantly improves the re-id accuracy compared to existing methods.

Introduction

Person Re-Identification (Re-ID) is the task of identifying individuals across non-overlapping camera views. While deep learning has revolutionized this field, the performance of these models is heavily dependent on the availability of large-scale labeled datasets. Manually labeling thousands of images across various camera angles and lighting conditions is an expensive and labor-intensive bottleneck.

Generative Adversarial Networks (GANs) have been employed to synthesize additional training data. However, assigning labels to these synthetic images is non-trivial. Previous methods, such as Label Smoothing Regularization for Outliers (LSRO), assign a uniform label across all identities. As the number of identities increases, this leads to "over-smoothness," where the probability assigned to any single identity becomes negligible, causing the model to treat synthetic data as noise rather than useful training signal.

The SLSR Framework

The authors introduce Sparse Label Smoothing Regularization (SLSR) to overcome the limitations of uniform smoothing. The framework operates on the principle that synthetic data should only provide regularization for visually similar identities.

1. Feature-Based Clustering

The training set is first processed using a pre-trained ResNet-50 to extract high-dimensional features. These features are then grouped using k-means clustering. This step ensures that identities sharing similar visual traits—such as clothing color, pose, or environmental lighting—are grouped together.

2. Cluster-Specific Generation

Rather than training a single global GAN, the authors train a Deep Convolutional GAN (DCGAN) for each cluster. This allows the generator to focus on the specific visual variations present within a cluster, producing more realistic and relevant synthetic images.

3. Sparse Label Assignment

The core innovation of SLSR is the sparse label distribution. Instead of spreading the label probability across the entire dataset, the generated image is assigned a uniform label only among the identities belonging to its parent cluster. For an identity $i$ in cluster $C$ with $p_c$ identities, the label $y_i$ is defined as $1/p_c$. For all identities outside the cluster, the label is $0$.

The SLSR pipeline: Clustering identities, generating samples from specific clusters, and assigning sparse labels based on cluster membership.

Experimental Results

The SLSR framework was evaluated on four major Re-ID benchmarks: Market-1501, CUHK03, DukeMTMC-ReID, and VIPeR.

  • Market-1501: SLSR achieved a Rank-1 accuracy of 89.16%, significantly outperforming the LSRO baseline of 83.97%.
  • CUHK03: The method showed a substantial improvement, reaching 91.03% Rank-1 accuracy, a jump of over 6% compared to previous GAN-based methods.
  • Robustness: The authors found that $K=3$ or $K=4$ clusters provided the best balance between regularization and discriminative power. Clustering works significantly better when performed on high-dimensional features extracted by a CNN rather than on raw pixels, as the CNN captures semantic similarity.

Conclusion

SLSR provides a robust solution to the over-smoothness problem in GAN-augmented Re-ID training. By leveraging the local structure of the data through clustering and sparse labeling, the framework ensures that synthetic images provide meaningful regularization. This approach significantly boosts model accuracy on large-scale datasets without requiring additional manual annotation.

Summary updated

Abstract

Person re-identification (re-id) is a cross-camera retrieval task which establishes a correspondence between images of a person from multiple cameras. Deep learning methods have been successfully applied to this problem and have achieved impressive results. However, these methods require a large amount of labeled training data. Currently, the labeled datasets in person re-id are limited in their scale and manual acquisition of such large-scale datasets from surveillance cameras is a tedious and labor-intensive task. In this paper, we propose a framework that performs intelligent data augmentation and assigns the partial smoothing label to generated data. Our approach first exploits the clustering property of existing person re-id datasets to create groups of similar objects that model cross-view variations. Each group is then used to generate realistic images through adversarial training. Our aim is to emphasize the feature similarity between generated samples and the original samples. Finally, we assign a non-uniform label distribution to the generated samples and define a regularized loss function for training. The proposed approach tackles two problems 1) how to efficiently use the generated data and 2) how to address the over-smoothness problem found in current regularization methods. The extensive experiments on four large-scale datasets show that our regularization method significantly improves the re-id accuracy compared to existing methods.
BibTeX
@inproceedings{ainam2019sparse,
  author = {Jean-Paul Ainam and Ke Qin and Guisong Liu and Guangchun Luo},
  title = {{Sparse Label Smoothing Regularization for Person Re-Identification}},
  booktitle = {IEEE Access},
  year = {2019},
  url = {https://openreview.net/forum?id=YPlzm8edgx},
}
Figure