Abstract
Most existing person re-identification methods are effective in short-term scenarios because of their appearance dependencies. However, these methods may fail in long-term scenarios where people might change their clothes. To this end, we propose an adversarial feature disentanglement network (AFD-Net) which contains intra-class reconstruction and inter-class adversary to disentangle the identity-related and identity-unrelated (clothing) features. For intra-class reconstruction, the person images with the same identity are represented and disentangled into identity and clothing features by two separate encoders, and further reconstructed into original images to reduce intra-class feature variations. For inter-class adversary, the disentangled features across different identities are exchanged and recombined to generate adversarial clothes-changing images for training, which makes the identity and clothing features more independent. Especially, to supervise these new generated clothes-changing images, a re-feeding strategy is designed to re-disentangle and reconstruct these new images for image-level self-supervision in the original image space and feature-level soft-supervision in the disentangled feature space. Moreover, we collect a challenging Market-Clothes dataset and a real-world PKU-Market-Reid dataset for evaluation. The results on one large-scale short-term dataset (Market-1501) and five long-term datasets (three public and two we proposed) confirm the superiority of our method against other state-of-the-art methods.
The Challenge of Long-Term Re-identification
Person Re-identification (Re-ID) is the task of matching a specific individual across different non-overlapping camera views. Most existing models are "short-term" systems that rely heavily on visual cues like the color of a shirt or the texture of a jacket. While effective for tracking someone through a mall in a single afternoon, these models fail in "long-term" scenarios where a person might be seen wearing different clothes on different days.
The core challenge is feature entanglement: in a standard neural network, the representation of a "person" is a mix of their physical build and their current outfit. To solve long-term Re-ID, we need a way to extract identity-related features that remain invariant even when the clothing changes.
Proposed Method: AFD-Net
The authors propose the Adversarial Feature Disentanglement Network (AFD-Net). The architecture relies on two separate encoders: $E_i$ for identity-related features (face, shape, hair) and $E_c$ for clothing features (color, texture, accessories).
The training process uses three main strategies:
- Intra-class Reconstruction: The model takes two images of the same person. It extracts the identity from one and the clothing from the other, then uses a generator ($G$) to reconstruct the original image. This forces the encoders to capture enough information to rebuild the person.
- Inter-class Adversary: The model takes images of two different people and swaps their features (Person A's identity + Person B's clothes). A discriminator ($D$) tries to distinguish these generated "clothes-changing" images from real ones. This adversarial process forces the identity encoder to ignore clothing cues and vice versa.
- Re-feeding Strategy: Since there is no "ground truth" image for what Person A would look like in Person B's clothes, the authors designed a re-feeding loop. The generated image is fed back into the encoders to see if the model can still extract the original identity and clothing features, providing a form of self-supervision.
Datasets and Evaluation
The study evaluates the model on six datasets, including three large-scale long-term benchmarks:
- Public Long-Term: Celeb-reID, Celeb-reID-light, and PRCC.
- Short-Term Benchmark: Market-1501 (to ensure the model still works in standard scenarios).
- New Benchmarks:
- Market-Clothes: A generated dataset based on Market-1501 where clothes are synthetically changed to create a massive long-term testing ground.
- PKU-Market-Reid: A real-world collected dataset from a supermarket with complex backgrounds, occlusions, and 3–10 sets of clothes per identity.
Key Results
- State-of-the-Art Performance: On the PRCC dataset (clothes-changing task), AFD-Net achieved a 42.8% Rank-1 accuracy, significantly outperforming previous methods like SPT+ASE (34.4%).
- Robustness: On their new Market-Clothes dataset, AFD-Net achieved 48.31% Rank-1, while standard models like DG-Net plummeted to 20.04%.
- Short-term Stability: The model maintained a high 94.2% Rank-1 on Market-1501, proving that disentangling features doesn't hurt performance when clothes stay the same.
- Visual Proof: Visualizations show that AFD-Net creates much tighter, more distinct clusters for different identities compared to previous state-of-the-art methods, effectively isolating identity from appearance.
Conclusion
AFD-Net demonstrates that by explicitly disentangling identity and clothing features through adversarial learning, Re-ID systems can become significantly more robust to temporal changes in appearance. While challenges remain—such as handling extreme variations in clothing volume (e.g., heavy coats)—this work provides a strong foundation for real-world, long-term surveillance and tracking applications.
Summary updated