Abstract
Person Re-Identification (ReID) in real-world scenarios is frequently hampered by occlusions from obstacles like vehicles or other pedestrians. While existing solutions often rely on complex auxiliary modules such as pose estimators or feature pyramids, this paper demonstrates that a streamlined approach using single-scale global representations can achieve state-of-the-art performance. By combining a novel data augmentation technique, reconstructive pooling, and a bounded exponential distance loss, the proposed method significantly improves accuracy on both occluded and standard ReID benchmarks while maintaining computational efficiency.
Introduction: The Challenge of Occlusion
Matching individuals across non-overlapping camera views is a fundamental task in computer vision. However, "Occluded ReID" remains a significant hurdle due to the unpredictable nature of obstacles in urban environments. Traditional high-performing models attempt to mitigate this by incorporating complex auxiliary components, such as pose estimators to identify body parts, or multi-scale feature pyramids.
These complex models face two primary issues: they are prone to failure when auxiliary modules make errors, and they often exhibit the "Clever Hans" effect—learning to identify the obstacle rather than the person. Furthermore, these specialized architectures frequently underperform on standard, non-occluded datasets.
Methodology: Robust Global Representations
The authors propose a simplified architecture based on a ResNet-50 backbone, enhanced by three key innovations designed to extract robust features from partially visible subjects.
1. Compound Batch Erasing (CBE)
To prevent the model from over-relying on specific body parts or background elements, the researchers introduced Compound Batch Erasing. This technique combines standard Random Erasing with Batch-constant Erasing (BcE), which masks the same spatial region across an entire training batch. This forces the network to seek out discriminative features in whatever body parts remain visible, regardless of the occlusion pattern.
2. Reconstructive Pooling (RP)
The backbone is augmented with Disentangled Non-Local (DNL) blocks to improve spatial focus. During training, a "reconstructive pooling" layer is employed to aggregate features specifically from non-occluded regions, ensuring the global representation is dominated by the person's identity rather than the occluding object.
3. Bounded Exponential Distance (BED) Loss
The "secret sauce" of the approach is the Bounded Exponential Distance (BED) loss. Unlike standard Triplet Loss, which can be unbounded and overly sensitive to outliers, the BED loss is exponentially sensitive to small distances. This forces the model to learn fine-grained distinctions between similar-looking individuals. Crucially, the loss is provably bounded, preventing noisy, heavily occluded samples from destabilizing the training process.
Datasets and the OPReID Benchmark
The model was evaluated on standard benchmarks including Occluded-Duke (O-Duke), Occlude-ReID (O-ReID), Market1501, and DukeMTMC-ReID. To further push the boundaries of the field, the authors introduced OPReID, a new large-scale dataset. OPReID contains over 74,000 images of 7,918 identities and features a wide variety of real-world occlusions (e.g., shopping carts, elevators, storage racks) captured across all four seasons.
Key Results and Insights
The proposed model demonstrates that simplicity can indeed outperform complexity:
- State-of-the-Art Performance: On the O-Duke dataset, the model achieved a 13.9% improvement in Top-1 accuracy over previous methods. On the new OPReID dataset, it outperformed baselines by 17-19%.
- Generalization: The model maintains high performance on non-occluded images, reaching 96.1% Rank-1 accuracy on Market1501.
- Efficiency: By utilizing a single backbone without auxiliary modules during inference, the system is significantly faster and more lightweight than its part-based or graph-based counterparts.
Conclusion
This research highlights that complex multi-stage pipelines are not strictly necessary for handling occlusions in ReID. A robust training regime—leveraging targeted data augmentation and a bounded loss function—allows a simple global feature representation to surpass more intricate models. Future work may focus on integrating the reconstructive logic directly into the inference stage to further enhance real-time performance.
Summary updated