Abstract
Person re-identification (Re-ID) has undergone a rapid development with the blooming of deep neural network. Most methods are very easily affected by target misalignment and background clutter in the training process. In this paper, we propose a simple yet effective feedforward attention network to address the two mentioned problems, in which a novel consistent attention regularizer and an improved triplet loss are designed to learn foreground attentive features for person Re-ID. Specifically, the consistent attention regularizer aims to keep the deduced foreground masks similar from the low-level, mid-level and high-level feature maps. As a result, the network will focus on the foreground regions at the lower layers, which is benefit to learn discriminative features from the foreground regions at the higher layers. Last but not least, the improved triplet loss is introduced to enhance the feature learning capability, which can jointly minimize the intra-class distance and maximize the inter-class distance in each triplet unit. Experimental results on the Market1501, DukeMTMC-reID and CUHK03 datasets have shown that our method outperforms most of the state-of-the-art approaches.
Introduction
Person Re-identification (Re-ID) is the task of matching the same individual across different, non-overlapping camera views. Despite significant progress, deep learning models for Re-ID often suffer from two types of "noise": background clutter and target misalignment. Background clutter occurs when a network mistakenly learns to identify a person based on environmental elements like trees or cars. Target misalignment arises from variations in pose and distance, causing body parts to occupy different relative positions in the image frame.
While attention mechanisms are commonly used to focus on foreground regions, they often fail at lower network layers where semantic information is sparse. This paper introduces a feedforward attention network that addresses these issues through a Consistent Attention Regularizer (CAR) and an improved triplet loss.
Methodology
Feedforward Attention Network Architecture
The proposed model is built upon a standard ResNet-50 backbone, enhanced with attention modules at multiple stages. The architecture integrates foreground masks that are applied to feature maps via element-wise products.
Consistent Attention Regularizer (CAR)
The core innovation is the Consistent Attention Regularizer (CAR). High-level layers naturally capture better semantic information about the person. CAR forces the attention masks (heatmaps) produced at low, mid, and high levels to remain consistent. By back-propagating this consistency requirement, high-quality foreground knowledge from deeper layers is distilled into shallower layers. This top-down guidance ensures the network maintains a consistent focus on the person throughout its depth.
The regularizer also includes an L1-norm sparsity term, which encourages the network to be selective and effectively suppress background noise by "turning off" irrelevant pixels.
Improved Triplet Loss
Standard triplet loss compares an anchor to one positive and one negative sample. The authors improve this by introducing a "center point" between the positive pair. This modification allows the model to more effectively minimize intra-class distance while maximizing inter-class distance, leading to more stable gradient updates and a more robust feature space.
Experimental Results
The model was evaluated on three major Re-ID benchmarks: Market1501, DukeMTMC-reID, and CUHK03.
- Market1501: Achieved a state-of-the-art 96.1% Top-1 accuracy.
- CUHK03 (Detected): Reached 93.2% Top-1 accuracy, significantly outperforming previous methods on automated detection boxes.
- Ablation Studies: Results showed that the Consistent Attention Regularizer alone consistently improved Top-1 accuracy by 2-3%, validating the effectiveness of the "consistent focus" training signal.
Conclusion
This work demonstrates that enforcing consistency across attention maps at different network depths is a powerful way to mitigate background noise and misalignment in person Re-ID. By combining this top-down guidance with an improved triplet loss, the proposed feedforward attention network achieves superior performance. Future work may explore incorporating temporal-spatial clues for video-based Re-ID and further refining hyperparameter sensitivity.
Summary updated