FD-GAN: Pose-guided Feature Distilling GAN for Robust Person Re-identification (1810.02936v2)

Abstract: Person re-identification (reID) is an important task that requires to retrieve a person's images from an image dataset, given one image of the person of interest. For learning robust person features, the pose variation of person images is one of the key challenges. Existing works targeting the problem either perform human alignment, or learn human-region-based representations. Extra pose information and computational cost is generally required for inference. To solve this issue, a Feature Distilling Generative Adversarial Network (FD-GAN) is proposed for learning identity-related and pose-unrelated representations. It is a novel framework based on a Siamese structure with multiple novel discriminators on human poses and identities. In addition to the discriminators, a novel same-pose loss is also integrated, which requires appearance of a same person's generated images to be similar. After learning pose-unrelated person features with pose guidance, no auxiliary pose information and additional computational cost is required during testing. Our proposed FD-GAN achieves state-of-the-art performance on three person reID datasets, which demonstrates that the effectiveness and robust feature distilling capability of the proposed FD-GAN.

Overview of FD-GAN: Pose-guided Feature Distilling GAN for Robust Person Re-identification

The paper presents a novel approach named Feature Distilling Generative Adversarial Network (FD-GAN) for addressing the problem of person re-identification (reID), which is the task of retrieving images of a person from a dataset based on a query image. The proposed method innovatively tackles the challenge of diverse human poses in reID tasks by learning identity-related features that are invariant to pose variations.

FD-GAN employs a Siamese network architecture consisting of an image encoder and generator paired with discriminators for identity and pose. This arrangement is designed to learn representations that are robust to pose variations without requiring additional pose information during testing. A distinctive feature of this approach is the incorporation of adversarial discriminators that help in distilling identity-related features by eliminating pose-related variances from encoded features.

Core Components and Technical Contributions

1. Siamese Network Structure: FD-GAN utilizes a Siamese network architecture to ensure robust learning of identity-related features across different human poses. Each branch of the Siamese network includes an image encoder and an image generator, supervised by a verification classifier to maintain identity feature consistency.
2. Generative Adversarial Framework: The authors employ a GAN framework where multiple discriminators enforce the feature learning process. The identity discriminator ensures the preserved person identity in the generated images, while the pose discriminator aids in mitigating pose-related information within the encoded features.
3. Novel Same-Pose Loss: To further refine the identity feature learning, a same-pose loss is introduced. This loss minimizes the difference between fake images across the two branches of the network generated under the same pose condition, promoting learning of pose-invariant features.
4. Elimination of Pose Information at Inference: Unlike traditional methods, FD-GAN does not require auxiliary pose information or additional computational overhead during testing, facilitating seamless reID performance deployment.
5. State-of-the-Art Performance: Experiments demonstrate that FD-GAN achieves superior performance across multiple reID datasets. The proposed framework not only surpasses existing methods in accuracy but also produces high-quality generated images.

Implications and Future Work

The implications of this research are significant for both practical and theoretical advancements in the field of person re-identification. Practically, the elimination of extra pose information and reduced computational costs during inference enhance the feasibility of deploying reID systems in dynamic environments. Theoretically, the integration of adversarial losses to disentangle identity-related features from pose-specific noise opens up new avenues for developing more refined and robust feature learning frameworks in computer vision.

Future work may focus on exploring different network architectures within the GAN framework to further enhance feature distillation capabilities. Additionally, the FD-GAN framework could be adapted to other challenging vision tasks where pose variation is a critical factor, such as activity recognition or biometric identification. Further investigation into the interpretability of learned features may also provide deeper insights into the model's decision-making process and improve the reliability and transparency of reID systems.

In conclusion, FD-GAN provides a compelling approach to overcoming the hurdles in person reID posed by diverse human poses. Through its sophisticated network structure and novel loss functions, it sets a new benchmark for continued research and development in identity-preserving machine learning models.