Learning Sparse and Identity-Preserved Hidden Attributes for Person Re-Identification

Zheng Wang, Junjun Jiang, Yang Wu, Mang Ye, Xiang Bai, Shin'ichi Satoh

Abstract

Person Re-Identification (Re-ID) is often hindered by the "granularity gap," where predefined semantic attributes are insufficient to distinguish between many individuals in large-scale datasets. This paper introduces Deep Hidden Attributes (DHA), an unsupervised approach to discover discriminative person features. By integrating a specialized ISO (Identity-preserving, Sparse, and Orthogonal) module into standard Re-ID networks, the proposed DHA-Net significantly improves matching accuracy across major benchmarks without requiring additional manual labels.

Introduction: The Granularity Gap

The task of matching the same individual across non-overlapping camera views typically relies on low-level visual features (color, texture) or high-level semantic attributes (gender, clothing). However, low-level features are sensitive to lighting and viewpoint changes, while semantic attributes are expensive to annotate and often lack the resolution needed for large populations.

The authors identify a critical "granularity gap": as the number of individuals in a dataset grows, the number of people sharing the same set of predefined attributes increases. Even with 15 semantic attributes, the average number of people sharing the same profile remains high, making unique identification difficult.

The granularity gap: average number of persons sharing common attributes vs. the number of predefined semantic attributes.

To address this, DHA-Net "mines" additional hidden attributes automatically from the data to provide the necessary discriminative power.

Proposed Method: DHA-Net and the ISO Module

The core of the proposed method is the ISO Module, an auto-encoder integrated into a baseline deep Re-ID network (e.g., ResNet-50). This module is governed by three specific constraints designed to learn effective hidden attributes:

  1. Orthogonal Generation (O): To prevent redundancy among hidden attributes, the model employs Singular Vector Decomposition (SVD) during training. This ensures that each learned attribute captures unique information.
  2. Identity-Preserving (I): Hidden attributes are constrained to be useful for identification. A cross-entropy loss ensures that the attribute profile for a specific person remains consistent across different images.
  3. Sparsity Constraint (S): An L1-like sparsity loss is applied to ensure that each person is defined by a unique, sparse "fingerprint" of active attributes, rather than a dense, non-discriminative feature vector.
The DHA-Net architecture, featuring the ISO Module and the multi-loss optimization framework.

DHA-Net optimizes four losses simultaneously: identification loss, semantic attribute prediction loss, visual reconstruction loss, and hidden attribute discrimination loss.

Experimental Evaluation

The researchers evaluated DHA-Net on major benchmarks, including Market-1501 and DukeMTMC-reID.

Key Results

  • Market-1501: Achieved 91.27% Rank-1 accuracy and 75.95% mAP, outperforming previous state-of-the-art methods like HACNN.
  • DukeMTMC-reID: Reached 81.32% Rank-1 accuracy.
  • Ablation Study: Adding the DHA module to a baseline ResNet-50 improved Rank-1 accuracy by approximately 3-4%, confirming that hidden attributes provide a significant performance boost over standard visual features alone.

Discussion and Insights

A key finding is that the network discovers "physical" meaning without human intervention. Visualization via Class Activation Maps showed that hidden attributes often focus on specific regions like backpack straps or shoe types—details that are highly discriminative but often omitted from standard semantic labeling schemas.

However, the method introduces some training complexity due to the SVD operation and the need to balance multiple loss functions. Future work may involve extending this "Identity-Preserved Hidden Attribute" framework to other domains, such as vehicle re-identification, where unique semantic markers are even more difficult to predefine.

Conclusion

DHA-Net effectively bridges the granularity gap in person Re-ID by automatically discovering sparse, orthogonal, and identity-preserving hidden attributes. This unsupervised feature mining approach provides a scalable way to enhance the discriminative power of deep Re-ID models without the burden of exhaustive manual annotation.

Summary updated

Abstract

Person re-identification (Re-ID) aims at matching person images captured in non-overlapping camera views. To represent person appearance, low-level visual features are sensitive to environmental changes, while high-level semantic attributes, such as “short-hair” or “long-hair”, are relatively stable. Hence, researches have started to design semantic attributes to reduce the visual ambiguity. However, to train a prediction model for semantic attributes, it requires plenty of annotations, which are hard to obtain in practical large-scale applications. To alleviate the reliance on annotation efforts, we propose to incrementally generate Deep Hidden Attribute (DHA) based on baseline deep network for newly uncovered annotations. In particular, we propose an auto-encoder model that can be plugged into any deep network to mine latent information in an unsupervised manner. To optimize the effectiveness of DHA, we reform the auto-encoder model with additional <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">orthogonal</i> generation module, along with <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">identity-preserving</i> and <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">sparsity</i> constraints. 1) <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">Orthogonally generating</i> : In order to make DHAs different from each other, Singular Vector Decomposition (SVD) is introduced to generate DHAs orthogonally. 2) <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">Identity-preserving constraint</i> : The generated DHAs should be distinct for telling different persons, so we associate DHAs with person identities. 3) <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">Sparsity constraint</i> : To enhance the discriminability of DHAs, we also introduce the sparsity constraint to restrict the number of effective DHAs for each person. Experiments conducted on public datasets have validated the effectiveness of the proposed network. On two large-scale datasets, <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">i.e.</i> , Market-1501 and DukeMTMC-reID, the proposed method outperforms the state-of-the-art methods.
BibTeX
@article{wang2020learning,
  author = {Zheng Wang and Junjun Jiang and Yang Wu and Mang Ye and Xiang Bai and Shin'ichi Satoh},
  title = {{Learning Sparse and Identity-Preserved Hidden Attributes for Person Re-Identification}},
  journal = {IEEE Trans. Image Process.},
  year = {2020},
  doi = {10.1109/tip.2019.2946975},
}
Figure