Proxy Task Learning For Cross-Domain Person Re-Identification

Abstract

Due to the lack of labels and the domain diversities, it is a challenge to study person re-identification in the crossdomain setting. An admirable method is to optimize the target model by assigning pseudo-labels for unlabeled samples through clustering. Usually, attributed to the domain gaps, the pre-trained source domain model cannot extract appropriate target domain features, which will dramatically affect the clustering performance and the accuracy of pseudo-labels. Extensive label noise will lead to sub-optimal solutions doubtlessly. To solve these problems, we propose an Anti-Noise Learning (ANL) approach, which contains two modules. The Feature Distribution Alignment (FDA) module is designed to gather the id-related samples and disperse idunrelated samples, through the camera-wise contrastive learning and adversarial adaptation. Creating a friendly cross-feature foundation for clustering that is to reduce clustering noise. Besides, the Reliable Sample Selection (RSS) module utilizes an Auxiliary Model to correct noisy labels and select reliable samples for the Main Model. In order to effectively utilize the outlier information generated by the clustering algorithm and RSS module, we train these samples at the instance-level. The experiments demonstrate that our proposed ANL framework can effectively reduce the domain conflicts and alleviate the influence of noisy samples, as well as superior performance compared with the state-of-the-art methods.

Receptive Multi-Granularity Representation for Person Re-Identification

“…When realizing the limitation of purely global feature learning, many attempts to local feature learning haven arisen. Some methods [7], [6], [33], [18], [34] refer to external clues of pose estimation or body part parsing to extract body part features of persons. [7], [6] utilize the structural part by pose estimation prediction to form relatively precise local region proposals for further representations.…”

“…[33] implements bilinear pooling between appearance features and body key point estimation features for part alignment on identity features. [18] stands on basic stripe-based feature learning and introduces structural pose clues for proper stripe splitting. [34] introduces 3D human body parsing to establish a semantic association between appearance and parsing results.…”

“…[10], [11], [3], [36] introduce a multi-level soft attention mechanism to enhance discrimination on part detail information of representations in spatial or channel dimensions. Stripe-based learning methods [4], [16], [20], [17], [18] commonly split the last feature map into several stripes as local representations on human bodies. [2], [37], [38], [17] rely on triplet or more complicated losses for distance metric learning.…”

“…However, this might not be a perfect solution for local representations due to some problems such as misalignment, which comes from the conflict between the assumption of within-part consistency and the infeasibility of perfect alignment of pedestrian images. Without external semantic guidance [33], [18], considering this principle and non-locality, feature representation is neither necessary nor reasonable. Here we explore another approach of dividing partitions from the view of feature selection.…”

“…Especially, recent local representation approaches [2], [3], [4], [5], [6], [7], [8], [9], [10], [11], [12], [13], [14] which extract features from local body parts by CNNs, have strongly proven robust to the environment variations. Among all these local representation methods, stripe-based feature learning [15], [2], [4], [16], [17], [18], [19], [20], [21] achieves the best tradeoff between performance and efficiency. Fig.…”