Unlabeled Samples Generated by GAN Improve the Person Re-identification Baseline in vitro (1701.07717v5)

Abstract: The main contribution of this paper is a simple semi-supervised pipeline that only uses the original training set without collecting extra data. It is challenging in 1) how to obtain more training data only from the training set and 2) how to use the newly generated data. In this work, the generative adversarial network (GAN) is used to generate unlabeled samples. We propose the label smoothing regularization for outliers (LSRO). This method assigns a uniform label distribution to the unlabeled images, which regularizes the supervised model and improves the baseline. We verify the proposed method on a practical problem: person re-identification (re-ID). This task aims to retrieve a query person from other cameras. We adopt the deep convolutional generative adversarial network (DCGAN) for sample generation, and a baseline convolutional neural network (CNN) for representation learning. Experiments show that adding the GAN-generated data effectively improves the discriminative ability of learned CNN embeddings. On three large-scale datasets, Market-1501, CUHK03 and DukeMTMC-reID, we obtain +4.37%, +1.6% and +2.46% improvement in rank-1 precision over the baseline CNN, respectively. We additionally apply the proposed method to fine-grained bird recognition and achieve a +0.6% improvement over a strong baseline. The code is available at https://github.com/layumi/Person-reID_GAN.

• The paper introduces LSRO, a novel method that assigns a uniform label distribution to GAN-generated samples to regularize CNN training.
• The proposed pipeline significantly boosts rank-1 accuracy and mAP, achieving up to +4.37% and +4.75% improvements on the Market-1501 dataset.
• This approach offers a cost-effective way to enhance person re-ID by leveraging unlabeled GAN data, pointing toward broader semi-supervised applications.

Unlabeled Samples Generated by GAN Improve the Person Re-identification Baseline in vitro

In the paper, the authors explore an innovative approach to enhancing person re-identification (re-ID) performance by integrating unlabeled samples generated using Generative Adversarial Networks (GANs). Specifically, their method aims to address the challenge of obtaining additional training data from an existing dataset and efficiently utilizing this newly generated data.

Methodological Framework

The research introduces a semi-supervised pipeline wherein GANs are employed to generate additional training samples from the original dataset. These samples, though unlabeled, are incorporated into the training of a Convolutional Neural Network (CNN) tasked with person re-ID tasks. The core innovation of this pipeline lies in the proposed Label Smoothing Regularization for Outliers (LSRO). LSRO assigns a uniform label distribution over the training classes to these GAN-generated samples. This technique prevents the network from making overly confident predictions for any specific class, thereby acting as a regularizer and mitigating the risk of overfitting.

Experimental Validation

The experiments carried out within the paper validate the efficacy of the proposed pipeline on the person re-ID task across three large-scale datasets: Market-1501, CUHK03, and DukeMTMC-reID. Utilizing a baseline Deep Convolutional GAN (DCGAN) for sample generation and a ResNet-50 model for CNN-based representation learning, the authors demonstrate significant improvements in discriminative capability when integrating GAN-generated data using LSRO.

• Market-1501: The method achieved a +4.37% increase in rank-1 accuracy compared to the baseline, reaching up to 78.06%. The mean Average Precision (mAP) also saw a notable improvement of +4.75% over the baseline.
• CUHK03: Modest improvements were observed, with +1.6% enhancement in rank-1 accuracy over the strong baseline.
• DukeMTMC-reID: The method attained a +2.46% increase in rank-1 accuracy coupled with a +2.14% rise in mAP.

Additionally, the method was extended to a fine-grained recognition task using the CUB-200-2011 dataset, achieving a +0.6% improvement over a strong baseline.

Comparative Analysis

To underscore the effectiveness of LSRO, the authors compared it against two alternative semi-supervised learning strategies: "All in one" and "Pseudo-labeling." Both methods demonstrated baseline improvements, yet LSRO outperformed them by approximately +1% to +2%. This suggests that LSRO's strategy of assuming GAN-generated samples as outliers with a uniform distribution more effectively regularizes the model.

Practical and Theoretical Implications

Practically, the ability to enhance performance by generating and utilizing GAN-created data without demanding additional real-world labeled data presents substantial utility in fields where data annotation is costly. Theoretically, the integration of LSRO offers a novel mechanism to fuse unlabeled data within supervised learning frameworks. This strategy presses future work towards exploring more sophisticated GAN models and investigating the application of such methodologies across other tasks requiring extensive labeled data.

Future Directions

Future research could delve into sophisticated GAN models for better visual quality in generated samples, potentially resulting in further improved regularization effects. Expanding the method's scope to other domains where labeled data is scarce but GAN-generated data is feasible could also yield beneficial insights into broadly applicable semi-supervised learning frameworks.

Conclusion

The presented approach demonstrates that leveraging GAN-generated images in a semi-supervised learning context can significantly bolster person re-ID performance. The proposed LSRO method serves as a robust regularizer within the training pipeline, enhancing the discriminative power of CNN embeddings across substantial datasets and tasks. This work contributes to the understanding and practical utilization of GANs in supervised learning, serving as a fundamental step towards more effective and scalable AI systems.