Abstract
Person re-identification (Re-ID) aims at retrieving an input person image from a set of images captured by multiple cameras. Although recent Re-ID methods have made great success, most of them extract features in terms of the attributes of clothing (e.g., color, texture). However, it is common for people to wear black clothes or be captured by surveillance systems in low light illumination, in which cases the attributes of the clothing are severely missing. We call this problem the Black Re-ID problem. To solve this problem, rather than relying on the clothing information, we propose to exploit head-shoulder features to assist person Re-ID. The head-shoulder adaptive attention network (HAA) is proposed to learn the head-shoulder feature and an innovative ensemble method is designed to enhance the generalization of our model. Given the input person image, the ensemble method would focus on the head-shoulder feature by assigning a larger weight if the individual insides the image is in black clothing. Due to the lack of a suitable benchmark dataset for studying the Black Re-ID problem, we also contribute the first Black-reID dataset, which contains 1274 identities in training set. Extensive evaluations on the Black-reID, Market1501 and DukeMTMC-reID datasets show that our model achieves the best result compared with the state-of-the-art Re-ID methods on both Black and conventional Re-ID problems. Furthermore, our method is also proved to be effective in dealing with person Re-ID in similar clothing.
Introduction: The "Black Re-ID" Problem
Modern person re-identification (Re-ID) systems are predominantly "clothing-centric," relying heavily on the color, texture, and patterns of a subject's attire. While effective in ideal conditions, this approach falters in two common real-world scenarios: the "Uniform Problem," where many individuals wear dark or black clothing, and low illumination conditions that cause colorful clothing to appear black.
The authors define this challenge as the Black Re-ID problem. When clothing attributes are uninformative, the system must pivot to more subtle, stable cues such as hairstyle, head shape, and accessories. These features are primarily concentrated in the head-shoulder region.
Proposed Method: Head-shoulder Adaptive Attention (HAA)
To address the limitations of global feature extraction, the authors propose the Head-shoulder Adaptive Attention (HAA) network. This two-stream architecture dynamically balances global full-body features with localized head-shoulder descriptors.
Head-shoulder Localization Layer (HLL)
A key innovation is the Head-shoulder Localization Layer (HLL). Unlike traditional methods that rely on computationally expensive pre-trained pose estimators, the HLL is a lightweight module inspired by Spatial Transformer Networks (STN). It learns to automatically "crop and zoom" into the head-shoulder region during the end-to-end training process, ensuring efficiency for real-time applications.
Head-shoulder Attention Network (HAN)
Once the region is localized, the Head-shoulder Attention Network (HAN) processes it by slicing the area into horizontal grids. It applies both channel and spatial attention to identify "micro-features," such as the presence of glasses or specific hair lengths, which provide high-entropy information even when the rest of the body is a silhouette.
Adaptive Attention Module
The ensemble's "brain" is the Adaptive Attention Module. It first predicts whether a subject is wearing black. If so, it automatically assigns a higher weight to the head-shoulder features. Conversely, for colorful attire, it relies more heavily on the global stream, ensuring robustness across diverse visual contexts.
The Black-reID Dataset
To facilitate research in this area, the authors introduced the Black-reID dataset, the first benchmark specifically designed for this problem. It contains 1,274 identities and is categorized into a "Black Group" focused on the core challenge and a "White Group" designed to test if the model's adaptive weighting generalizes to other low-variance uniform colors. The model was also evaluated on standard benchmarks like Market1501 and DukeMTMC-reID to confirm its general-purpose effectiveness.
Experimental Results
The HAA model achieved state-of-the-art (SOTA) performance across multiple datasets:
- Black-reID: Achieved 83.8% mAP and 91.0% Rank-1 accuracy, significantly outperforming previous SOTA models like MGN.
- Market1501 & DukeMTMC-reID: Showed consistent improvements (2.6% and 2.0% mAP respectively), proving that focusing on head-shoulder features benefits general Re-ID tasks as well.
Qualitative analysis confirms that while standard models are often confused by multiple individuals in similar dark jackets, HAA successfully distinguishes them by focusing on stable physical attributes like shoulder width and hair style.
Conclusion
The HAA network effectively mitigates the "Black Re-ID" problem by shifting focus from volatile clothing attributes to stable head-shoulder features. By utilizing a lightweight localization layer and dynamic feature weighting, the proposed method offers a robust and efficient solution for real-world surveillance challenges where clothing information is often missing or uninformative.
Summary updated