Abstract
In person re-IDentification (re-ID) task, the learning of part-level features benefits from fine-grained information. To facilitate part alignment, which is a prerequisite for learning part-level features, a popular approach is to detect semantic parts with the use of human parsing or pose estimation. Such methods of semantic partition do offer cues to good part alignment but are prone to noisy part detection, especially when they are employed in an off-the-shelf manner. In response, this paper proposes a novel part feature learning method for re-ID, that suppresses the impact of noisy semantic part detection through Supervised Non-local Similarity (SNS) learning. Given several detected semantic parts, SNS first locates their center points on the convolutional feature maps for use as a set of anchors and then evaluates the similarity values between these anchors and each pixel on the feature maps. The non-local similarity learning is supervised such that: each anchor should be similar to itself and simultaneously dissimilar to any other anchors, thus yielding the SNS. Finally, each anchor absorbs features from all of the similar pixels on the convolutional feature maps to generate a corresponding part feature (SNS feature). We evaluate our method with extensive experiments conducted under both holistic and partial re-ID scenarios. Experimental results confirm that SNS consistently improves re-ID accuracy using human parsing or pose estimation, and that our results are on par with state-of-the-art methods.
Introduction: The Challenge of Noisy Part Detection
Person re-identification (re-ID)—the task of matching the same individual across different camera views—is a cornerstone of modern surveillance and search systems. To achieve high accuracy, models often look at "part-level" features (e.g., the head, torso, or legs) rather than just a global image.
The standard way to find these parts is to use off-the-shelf human parsing or pose estimation models. However, there is a significant domain gap: a parser trained on high-quality fashion photos often struggles with the low-resolution, blurry, or occluded images typical of re-ID datasets. This "noisy" detection leads to part extraction errors, which ultimately degrades re-ID performance. This research introduces a mechanism that leverages the cues provided by these parsers without being strictly bound by their inaccurate masks.
The SNS Framework
The proposed Supervised Non-local Similarity (SNS) framework addresses the noise in semantic part detection through a three-stage process:
- Anchor Localization: Instead of using the entire mask from a human parser, the model identifies the geometric center of each detected part (head, upper body, lower body, feet) on the convolutional feature map. These centers serve as anchors.
- Supervised Similarity Learning: The model calculates the similarity between these anchors and every other pixel on the feature map. Crucially, this is supervised. An SNS loss forces each anchor to be highly similar to itself and its immediate region, while remaining dissimilar to anchors representing other body parts.
- Feature Absorption: Each anchor "absorbs" features from all pixels across the map, weighted by their similarity. If a pixel in the corner of the image looks like "head" material, the head anchor will pull information from it. This allows the model to "reconstruct" a clean part feature even if the initial parser only caught a glimpse of the head.
Key Insights and Visualizations
A core observation of this work is that centers are more stable than masks. Even when a human parser produces a blurry or truncated mask, the geometric center usually still lands within the correct body region. By treating these centers as starting points (anchors) rather than absolute boundaries, the model gains robustness.
Furthermore, while standard non-local blocks are typically unsupervised, the addition of a specific SNS loss forces the network to learn semantically meaningful relationships. This ensures that the "absorption" process is grounded in actual body part characteristics, effectively cleaning up the noisy input from the parser.
Experimental Evaluation
The SNS method was evaluated on four benchmarks covering both holistic and partial re-ID scenarios: Market-1501, DukeMTMC-reID, Partial-REID, and Partial-iLIDS.
Key Results
- Market-1501: The method achieved a 91.5% rank-1 accuracy and 78.6% mAP, a significant jump over the standard IDE baseline (87.3% rank-1).
- Robustness: When compared to a baseline that directly uses parsing masks (H-baseline), SNS improved rank-1 accuracy by up to 2.7%, proving its ability to mitigate noisy detections.
- Partial re-ID: On the Partial-REID dataset, SNS achieved 52.4% rank-1 accuracy, significantly outperforming the popular PCB (Part-based Convolutional Baseline) which hit 34.6%. This highlights the advantage of semantic parts over rigid horizontal stripes when the person is occluded. For partial re-ID, the model also utilizes a "visibility score" to ignore missing parts during comparison.
Conclusion
The SNS mechanism successfully mitigates the impact of noisy semantic part detection by using supervised non-local similarity to refine part features. While the method shows strong performance, it remains dependent on the initial anchor placement; extreme parser failure can still lead to misalignment. Future work may explore iterative anchor localization to further enhance robustness.
Summary updated