AD-Cluster: Augmented Discriminative Clustering for Domain Adaptive Person Re-identification

Abstract: Domain adaptive person re-identification (re-ID) is a challenging task, especially when person identities in target domains are unknown. Existing methods attempt to address this challenge by transferring image styles or aligning feature distributions across domains, whereas the rich unlabeled samples in target domains are not sufficiently exploited. This paper presents a novel augmented discriminative clustering (AD-Cluster) technique that estimates and augments person clusters in target domains and enforces the discrimination ability of re-ID models with the augmented clusters. AD-Cluster is trained by iterative density-based clustering, adaptive sample augmentation, and discriminative feature learning. It learns an image generator and a feature encoder which aim to maximize the intra-cluster diversity in the sample space and minimize the intra-cluster distance in the feature space in an adversarial min-max manner. Finally, AD-Cluster increases the diversity of sample clusters and improves the discrimination capability of re-ID models greatly. Extensive experiments over Market-1501 and DukeMTMC-reID show that AD-Cluster outperforms the state-of-the-art with large margins.

• The paper introduces AD-Cluster, a novel augmented discriminative clustering method that enhances domain adaptive person re-identification by leveraging unlabelled target data through iterative clustering, adaptive sample augmentation, and discriminative feature learning.
• Experimental results show significant performance improvements over state-of-the-art methods on benchmark datasets like Market-1501 and DukeMTMC-reID, demonstrating increased rank-1 accuracy and mAP in cross-domain transfer scenarios.
• AD-Cluster offers a practical contribution to unsupervised domain adaptation by efficiently exploiting unlabelled data and shows theoretical promise for broader application in general recognition and object detection tasks.

Overview of AD-Cluster for Domain Adaptive Person Re-identification

The paper presents a novel approach, named AD-Cluster, which tackles the problem of domain adaptive person re-identification (re-ID). This task is particularly challenging due to the fact that person identities in target domains are typically unlabeled, and current methods depend heavily on transferring image styles or feature distribution alignment, often neglecting the abundance of unlabelled target samples.

AD-Cluster introduces an augmented discriminative clustering method to enhance the re-identification process by estimating and augmenting person clusters in the target domain. It enforces the discriminative capability of re-ID models through an innovative approach combining iterative density-based clustering, adaptive sample augmentation, and discriminative feature learning. The system utilizes an image generator and a feature encoder to maximize intra-cluster diversity in the sample space while minimizing intra-cluster distances in the feature space using an adversarial min-max strategy. This method effectively enhances both the diversity of sample clusters and the discriminative power of re-ID models.

Methodology

AD-Cluster's approach is multidimensional:

1. Density-Based Clustering: The system begins by predicting sample clusters in the target domain. It uses a model pretrained on the source domain to extract these features.
2. Adaptive Sample Augmentation: Employs a Generative Adversarial Network (GAN) to translate images across different camera styles, thereby augmenting the diversity of training samples while maintaining identity consistency.
3. Discriminative Feature Learning: The feature encoder is trained iteratively to maximize inter-cluster distances and minimize intra-cluster distances, promoting a more discriminative feature representation.

These elements create a framework where an image generator and a feature encoder operate adversarially. This adversarial min-max training optimizes both generated cluster labels and re-ID model performance.

Experimental Results

The experiments conducted as part of this research demonstrate significant performance improvement over existing state-of-the-art methods on the Market-1501 and DukeMTMC-reID datasets. AD-Cluster shows a considerable margin of improvement, with a rank-1 accuracy of 86.7% and mAP of 68.3% for the DukeMTMC-reID to Market-1501 transfer scenario, and 72.6% and 54.1% respectively for the Market-1501 to DukeMTMC-reID.

Implications and Future Directions

Practically, AD-Cluster contributes significantly to the domain of unsupervised domain adaptation for person re-ID. It efficiently exploits unlabelled target data, thus pushing the boundaries in situations where labeled data is scarce or expensive to obtain. Theoretically, it proves the prowess of adversarial training in enhancing model discrimination through augmented and diverse data samples.

Looking forward, AD-Cluster might be extended or adapted to other domain adaptation challenges broader than re-ID, such as general recognition or object detection. Further exploration could involve integrating more sophisticated augmentation techniques or leveraging additional metadata (e.g., timestamps, GPS) to even better exploit the unlabeled samples and refine the unsupervised learning process. The introduction of such novel methodologies would undoubtedly drive further advancements in machine learning's ability to generalize across domains.