Abstract
Unsupervised domain adaptation on person re-identification (re-ID), which adapts the model trained on source dataset to the target dataset, has drawn increasing attention over the past few years. It is more practical than the traditional supervised methods when applied in the real-world scenarios since they require a huge number of manual annotations in a specific domain, which is unrealistic and even under personal privacy concerns. Currently, pseudo label-based method is one of the most promising solutions in this area. However, in such methods, pseudo label noise is ignored and remains a huge challenge hindering further performance improvements. To solve this problem, this paper proposes a novel unsupervised domain adaptation re-ID framework named Noise Resistible Network (NRNet), which mainly consists of two dual-stream networks. For one thing, during pseudo label generation, NRNet utilizes one dual-stream network, denoted as clustering network, to generate discriminative features in the unseen domain for further clustering, reducing the pseudo label noise. For another, to avoid the problem of close loop noise amplification in conventional methods, the other dual-stream network named temporally average network is constructed outside the clustering loop to learn how to identify the images of the same person. In addition, two dual-stream networks are designed with a guiding mechanism, which allows the shallow network to learn more representative feature embedding from the deep network. Extensive experimental results on two widely-used benchmark datasets, i.e., Market-1501 and DukeMTMC-reID demonstrate that our proposed NRNet outperforms the state-of-the-art methods.
Introduction
Person re-identification (re-ID) is the task of matching images of the same person across different, non-overlapping camera views. While supervised learning has made massive strides in this field, it requires thousands of manual annotations—a process that is expensive, time-consuming, and raises significant privacy concerns. Unsupervised Domain Adaptation (UDA) offers a solution by training a model on a labeled "source" dataset and adapting it to an unlabeled "target" dataset.
The Challenge of Pseudo-Label Noise
Most modern UDA methods use pseudo-labeling, where the model clusters unlabeled target images and treats those clusters as ground truth. However, this approach faces a major hurdle: Pseudo Label Noise. If the model’s features aren't discriminative enough, it groups different people together or splits the same person into different clusters.
This noise leads to Close-Loop Noise Amplification. When a model is trained on its own noisy pseudo-labels, it becomes more confident in its mistakes, leading to a "garbage in, garbage out" spiral that degrades performance over time.
The NRNet Framework
The authors propose the Noise Resistible Network (NRNet), which breaks the noise cycle using three main components designed to stabilize the iterative training process.
Key Innovations
- Dual-Stream Clustering Network: Instead of one model, NRNet uses two different architectures (ResNet-50 and ResNet-152). By fusing the features from these heterogeneous networks, the system gains a more "complete" view of the person, leading to more robust and accurate clusters.
- Temporally Average Network: To solve the loop amplification problem, the authors construct a separate network outside the main training loop. This network is updated using a moving average of the clustering network's parameters. This "temporal smoothing" makes the model fault-tolerant; if one iteration produces bad labels, the averaged model isn't immediately derailed.
- Guiding Mechanism: The deeper, more powerful ResNet-152 "guides" the shallower ResNet-50. This allows the smaller model (which is often preferred for deployment due to speed) to learn more representative feature embeddings.
Experimental Results
NRNet was evaluated on two major benchmarks: Market-1501 (32,668 images) and DukeMTMC-reID (36,411 images). The framework outperformed existing state-of-the-art (SOTA) methods across the board:
- Duke → Market-1501: Achieved a 91.1% Rank-1 accuracy and 78.6% mAP (mean Average Precision), outperforming the previous SOTA (MMT-500) by 3.4%.
- Market-1501 → Duke: Achieved a 79.8% Rank-1 accuracy and 67.6% mAP.
Qualitative analysis using t-SNE visualizations showed that NRNet creates much tighter, more distinct clusters for different individuals compared to models that are simply transferred without adaptation.
Conclusion
NRNet effectively resists pseudo-label noise by leveraging architectural heterogeneity and temporal averaging. While a gap still exists between unsupervised adaptation and fully supervised performance, NRNet provides a robust and efficient framework for real-world person re-identification where manual labeling is unfeasible.
Summary updated