Abstract
Previous Person Re-Identification (Re-ID) models aim to focus on the most discriminative region of an image, while its performance may be compromised when that region is missing caused by camera viewpoint changes or occlusion. To solve this issue, we propose a novel model named Hierarchical Bi-directional Feature Perception Network (HBFP-Net) to correlate multi-level information and reinforce each other. First, the correlation maps of cross-level feature-pairs are modeled via low-rank bilinear pooling. Then, based on the correlation maps, Bi-directional Feature Perception (BFP) module is employed to enrich the attention regions of high-level feature, and to learn abstract and specific information in low-level feature. And then, we propose a novel end-to-end hierarchical network which integrates multi-level augmented features and inputs the augmented low- and middle-level features to following layers to retrain a new powerful network. What's more, we propose a novel trainable generalized pooling, which can dynamically select any value of all locations in feature maps to be activated. Extensive experiments implemented on the mainstream evaluation datasets including Market-1501, CUHK03 and DukeMTMC-ReID show that our method outperforms the recent SOTA Re-ID models.
1. Introduction
Person Re-Identification (Re-ID) is the task of identifying the same individual across different non-overlapping camera views. While deep learning has significantly advanced the field, many existing models suffer from "salience bias." They tend to fixate on the most obvious features—such as a bright backpack—while ignoring "non-salient" but discriminative details like shoe types or clothing patterns. This bias leads to failure when salient features are occluded or when different individuals share similar prominent accessories.
To address this, we propose the Hierarchical Bi-directional Feature Perception Network (HBFP-Net). The core idea is to enable "bi-directional communication" between low-level textures and high-level semantic features, forcing the model to capture a more holistic representation of the person.
2. Proposed Method
The HBFP-Net architecture, shown in Figure 1, builds upon a ResNet-50 backbone and introduces two key components: the Bi-directional Feature Perception (BFP) module and Generalized Pooling (GeP).
2.1 Bi-directional Feature Perception (BFP)
Standard convolutional networks extract features hierarchically, but information flow is typically one-way. The BFP module utilizes low-rank bilinear pooling to calculate correlation maps between different feature levels (low, middle, and high). This allows for mutual reinforcement:
- High-to-Low: Semantic context from deeper layers helps shallower layers filter out background noise and focus on the human body.
- Low-to-High: Fine-grained details (textures, edges) are injected into deeper layers, expanding the model's attention from a single "hotspot" to the entire body.
2.2 Generalized Pooling (GeP)
Traditional pooling methods like Max Pooling or Average Pooling can either lose subtle information or be diluted by noise. We introduce Generalized Pooling (GeP), which uses a trainable threshold ($\lambda$). GeP dynamically averages only the values that exceed this threshold, allowing the model to capture "gradual cues"—useful features that are discriminative but not necessarily the strongest signals in the feature map.
3. Experimental Results
HBFP-Net was evaluated on three major benchmarks: Market-1501, DukeMTMC-ReID, and CUHK03-NP.
3.1 Performance Comparison
The model achieved state-of-the-art results across all datasets:
- Market-1501: 95.8% Rank-1 accuracy and 89.8% mAP.
- DukeMTMC-ReID: 89.5% Rank-1 and 80.2% mAP.
- CUHK03-NP (Detected): 80.0% Rank-1, significantly outperforming previous part-based and attention-based models.
3.2 Visualization of Attention
To understand why HBFP-Net performs better, we visualized the attention maps using Grad-CAM. As shown in Figure 2, standard models often focus narrowly on a single salient feature like a backpack. In contrast, HBFP-Net spreads its attention across the entire body, including legs and feet. This holistic view makes the model much more robust to occlusions.
4. Conclusion
HBFP-Net addresses the salience bias in person Re-ID by correlating multi-level features through a bi-directional perception mechanism. By integrating fine-grained details with semantic context and employing a trainable generalized pooling layer, the model learns robust, holistic representations. Experimental results confirm that this approach significantly improves performance and robustness against common challenges like occlusion and viewpoint changes.
Summary updated