Abstract
In this paper, we propose a Resetting-Label Network based on Fast Group Loss for Person Re-Identification (RLFGL-ReID). The major challenge of Re-ID lies in how to preserve the similarity of the same person against large variations caused by complex background, different illuminations, and various view angles while discriminating different individuals. To address the above-mentioned problems, we propose the RLFGL-ReID that includes resetting-label (RL) and fast group loss (FGL). Two main contributions of our network are as follows. First, a new method, the resetting-label method, which resets the ID labels, is proposed for the Re-ID network. Resetting the ID labels of each pedestrian is beneficial in maximizing the inter-group distances between each people, achieving better performance on classifying different individuals. Second, a fast group loss, i.e., an advanced version of variance group loss (VGL), is proposed to simplify the training process and accelerate the loss computation. By doing so, the network can eliminate the restriction of inputting the whole group of data when training the network. To confirm the effectiveness of our method, we extensively conduct our method on several widely used person Re-ID benchmark datasets. As the result shows, our method achieves rank@1 accuracy of 98.38% on CUHK03, 95.46% on Market1501, and 91.2% on DukeMTMC-reID, outperforming the state-of-the-art methods and confirming the advantage of our method.
The Challenge of Person Re-Identification
Person Re-Identification (Re-ID) is the task of matching the same individual across different, non-overlapping camera views. It is a core component of modern surveillance and multi-camera tracking systems. However, it remains incredibly difficult due to intra-class variation (the same person looks different due to lighting, pose, or occlusion) and inter-class similarity (different people wearing similar clothes).
Current deep learning approaches often face two specific bottlenecks:
- Arbitrary Labeling: Standard datasets assign ID labels (e.g., Person 1, Person 2) arbitrarily. These fixed labels may not align with the natural "shortest path" for a network to learn distinct features, potentially slowing down convergence.
- Loss Function Complexity: Advanced losses like Variance Group Loss (VGL) are effective at minimizing intra-group distance but are computationally expensive and require specific, restrictive data batching (inputting all images of a single ID at once).
The RLFGL-ReID Framework
The proposed Resetting-Label Network based on Fast Group Loss (RLFGL-ReID) introduces three main innovations to overcome these hurdles:
Resetting-Label (RL)
This is the paper's most novel contribution. Instead of sticking to the dataset's original labels, the authors periodically extract average features for each person, sort them, and re-assign ID labels. This "rearrangement" ensures that the inter-group distances between people are maximized in a way that aligns with the network's current feature space, making the classification task easier for the model.
Fast Group Loss (FGL)
FGL is an evolution of Variance Group Loss. While VGL requires a whole group of images for a single ID to calculate a loss, FGL uses a mathematical formulation that allows for random image input. This simplifies the training pipeline, reduces memory overhead, and accelerates computation while still effectively minimizing intra-group variance.
3D-Transpose Feature Extraction (3DTFE)
Standard convolutions often struggle to capture the full spatial context of a pedestrian. 3DTFE transposes the image tensor to process "bars" in the width direction. Since pedestrians are vertically oriented, this method better captures horizontal spatial features (like patterns on a shirt or a bag strap) that are highly discriminative.
Experimental Results
The model was evaluated on three gold-standard Re-ID benchmarks: CUHK03, Market-1501, and DukeMTMC-reID. The RLFGL-ReID model outperformed several state-of-the-art (SOTA) methods across all metrics:
- Accuracy: Achieved 98.38% Rank@1 on CUHK03, 95.46% on Market-1501, and 91.2% on DukeMTMC-reID.
- Speed: In head-to-head tests on CUHK03, the Fast Group Loss (FGL) completed 100 epochs in 10,575 seconds, significantly faster than VGL (13,910s) and Triplet Loss (14,272s).
- Ablation Studies: Adding the Resetting-Label (RL) module to the baseline consistently provided a ~0.4% to 0.8% boost in Rank@1 accuracy across all datasets, proving that dynamic label assignment is a powerful tool for feature optimization.
Conclusion
The RLFGL-ReID framework demonstrates that ID labels in a classification task should not be treated as sacred constraints. By treating them as dynamic variables that can be re-mapped to fit the learned feature distribution, and by simplifying group-based distance metrics through FGL, the authors achieve superior accuracy and training efficiency. Future work may involve integrating the label-resetting process into a single, end-to-end continuous training loop.
Summary updated