Cloth-Invariant Feature Learning for Person Re-Identification

29 concepts18 papers1 proposal

Traditional person re-identification (Re-ID) models suffer from a fundamental shortcut: they are highly effective clothing detectors, but poor identity recognizers. As Igor Barros Barbosa et al. observed in Looking beyond appearances: Synthetic training data for deep CNNs in re-identification, deep neural networks heavily rely on visual appearance—specifically clothing color and texture—as a proxy for identity.

This reliance causes a catastrophic drop in accuracy in long-term surveillance scenarios where individuals change their outfits. When clothes change, the intra-identity variance (how much a single person's appearance changes) suddenly dwarfs the inter-identity variance (the difference between two distinct people wearing similar clothes), as noted by Peixian Hong et al. in Fine-Grained Shape-Appearance Mutual Learning for Cloth-Changing Person Re-Identification. To build robust systems, researchers have shifted toward cloth-invariant feature learning, developing mechanisms to actively strip away transient clothing cues while preserving persistent biometric signals.

Identity-Appearance Disentanglement

Instead of simply hoping a network ignores clothing, a more effective strategy is to explicitly separate the latent space into distinct identity and clothing components. Yuming Yan et al. demonstrate in Weakening the Influence of Clothing: Universal Clothing Attribute Disentanglement for Person Re-Identification that actively disentangling these features prevents clothing attributes from acting as a disturbance.

By isolating clothing features into their own subspace, the identity subspace is forced to rely on traits that do not change. For instance, Wuxuan Shi et al. propose Structural Identity Decoupling in Towards Realistic Lifelong Re-identification: Identity Recurrence with Changing Clothes to separate permanent body structures from transient outfits. To guide this process without requiring expensive frame-level clothing annotations, Haoxuan Xu and Guanglin Niu introduce a dual-stream architecture in Identity-aware Feature Decoupling Learning for Clothing-change Person Re-identification, transferring spatial knowledge from a clothing-masked attention stream to a standard RGB stream. Similarly, the authors of Colors See Colors Ignore: Clothes Changing ReID with Color Disentanglement treat color as a direct proxy for appearance bias, explicitly disentangling it to achieve cloth-invariance using only standard RGB data.

Adversarial Clothing Suppression

How do we mathematically guarantee that an identity representation contains no clothing information? The field leverages adversarial training to actively suppress clothing signals.

In Clothes-Changing Person Re-identification with RGB Modality Only, Hong Chang et al. introduce a Clothes-based Adversarial Loss (CAL). The mechanism operates as a minimax game: an auxiliary classifier is trained to identify which outfit a person is wearing, while the feature extractor is optimized to maximize the classifier's confusion. Because the feature extractor must actively prevent the classifier from differentiating between outfits worn by the same individual, it is forced to discard transient clothing features and focus exclusively on permanent physical traits like body shape or hairstyle.

Shape and Pose as Invariant Proxies

When surface-level textures and colors are stripped away, the model must find alternative, stable biometric anchors. Body shape and skeletal structure serve as the primary proxies for identity.

Peixian Hong et al. leverage this in Fine-Grained Shape-Appearance Mutual Learning for Cloth-Changing Person Re-Identification via a two-stream mutual learning framework that explicitly teaches an appearance model to capture subtle, identity-linked body shape cues. To make these structural features more explicit, Quoc-Huy Trinh et al. introduce PGDS: Pose-Guidance Deep Supervision for Mitigating Clothes-Changing in Person Re-Identification, using a frozen pose estimation model to inject structural topology directly into the human encoder. Taking this further into unconstrained environments, Haidong Zhu et al. present ShARc in ShARc: Shape and Appearance Recognition for Person Identification In-the-wild, a multimodal framework that fuses 3D body shape, pose, and appearance to bypass clothing volatility entirely.

Synthesizing Intra-Identity Variance

A major bottleneck in training cloth-invariant models is data scarcity. Standard Re-ID datasets rarely feature the same individual in multiple outfits, making it difficult for models to learn what changes versus what stays constant. As Peng Zhang et al. highlight in Cloth-changing person re-identification: a summary, manually labeling such datasets is highly resource-intensive.

To circumvent this, researchers use Generative Adversarial Networks (GANs) to augment datasets with synthetic clothing variations. In Color-Sensitive Person Re-Identification, Guan'an Wang et al. introduce a Color Translation GAN (CTGAN) to transfer different colors onto a person's existing clothes. This allows the network to learn which color variations belong to the same identity. Expanding on this concept, Zhengxu Yu et al. develop an AS-GAN in Apparel-Invariant Feature Learning for Person Re-Identification to generate entirely new clothing styles and textures for a single individual, artificially expanding the intra-identity variance during training.

Forcing Invariance via Hard Positive Contrastive Learning

With synthetic or real cloth-changing data available, the final piece of the puzzle is designing loss functions that penalize the model for being fooled by clothing changes. Standard contrastive learning treats all positive pairs (images of the same person) equally. However, matching the same person in two different outfits is a significantly harder task than matching them in the same outfit.

To address this, Haoxuan Xu and Guanglin Niu employ a specialized contrastive loss function in Identity-aware Feature Decoupling Learning for Clothing-change Person Re-identification that explicitly weights "hard positive" pairs—instances where the same individual appears in entirely different clothing. By heavily penalizing large feature distances between these hard positives, the loss function forces the network to look past superficial color and texture changes, ensuring that the final embedding space groups individuals by their underlying identity rather than their attire.

Key Foundational Work

Looking beyond appearances: Synthetic training data for deep CNNs in re-identification (2018) — contributed 1 concept elements to this direction.
Color-Sensitive Person Re-Identification (2019) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_cloth_changing_person_re_identification_70adcd3a,
  title = {cloth-changing person re-identification},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/cloth-changing-person-re-identification/rdr_70adcd3acac747b080a390ed4e05963c},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_cloth_changing_person_re_identification_70adcd3a_v1,
  title = {cloth-changing person re-identification},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/cloth-changing-person-re-identification/rdr_70adcd3acac747b080a390ed4e05963c/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake