Abstract
Cloth-changing person re-identification (CC-ReID) presents a significant challenge for traditional surveillance systems that rely heavily on visual appearance and clothing textures. Existing solutions often utilize complex 3D body modeling or silhouettes, which can be computationally expensive or discard valuable stable color information. This paper introduces Masked Attribute Description Embedding (MADE), a novel framework that leverages editable attribute descriptions to identify individuals across clothing changes. By explicitly masking cloth-sensitive attributes (e.g., shirt color) while retaining stable personal traits (e.g., age, gender, hair), MADE achieves state-of-the-art performance across multiple benchmarks, including PRCC, LTCC, and LaST.
Introduction
Standard Person Re-identification (Re-ID) systems typically assume that a person's clothing remains constant between different camera views. However, in long-term surveillance or "in-the-wild" scenarios, individuals frequently change their outfits. This causes traditional models to fail because the dominant visual signal—the clothes—is no longer a reliable identifier.
Previous attempts to address CC-ReID focused on "cloth-irrelevant" biological features such as silhouettes, sketches, or 3D body shapes. While these methods mitigate the clothing bias, they often suffer from high computational complexity, instability in generative training (e.g., GANs), and significant information loss. For instance, silhouettes discard useful color information that remains consistent, such as hair or shoe color. MADE addresses these limitations by using a simpler, more flexible "editable" attribute approach.
Methodology
The MADE framework consists of three primary components designed to prioritize stable personal features over transient fashion choices.
Description Extraction and Mask (DEM)
The system first utilizes a pre-trained attribute detection model (SOLIDER) to extract a vector of 105 pedestrian attributes. These include traits like "Male," "Long Hair," and "Red Upper Color." To force the model to ignore clothing, the authors implement a manual masking operation where four specific categories—upper body type, upper body color, lower body type, and lower body color—are set to zero. This results in a "cloth-irrelevant" attribute vector.
Attribute Embedding and Fusion
Instead of employing a heavy text encoder, MADE uses a simple Linear Projection to map the binary attribute vector into the same dimensionality as the image features. These embeddings are then injected into a Vision Transformer (EVA02-large) at multiple stages (Stage 2 and Stage 3). This allows the model to integrate high-level semantic descriptions with visual features.
Multi-level Feature Fusion
The final representation is created by fusing low-level and high-level features from different stages of the Transformer using a Conv1D layer. This ensures the model captures both fine-grained visual details and robust semantic attributes.
Key Insights
- Attributes as Soft Biometrics: The core innovation is treating textual attributes as editable "soft biometrics." By zeroing out specific bits in a vector, the model is explicitly instructed on which features to ignore, a process far more efficient than 3D reconstruction.
- Robustness to Noise: Interestingly, adding approximately 10% random noise to attribute labels during training improved performance. This suggests the model learns to treat attributes as helpful hints rather than absolute truths, making it resilient to errors from the initial attribute extractor.
- Preserving Stable Color: Unlike silhouette-based methods that go entirely grayscale, MADE retains color information for hair and shoes, which are often consistent even when a person changes their primary outfit.
Experimental Results
MADE was evaluated on four major CC-ReID benchmarks, consistently outperforming previous state-of-the-art (SOTA) methods:
| Dataset | Rank-1 Accuracy | Improvement over previous SOTA |
|---|---|---|
| PRCC | 64.3% | +~6.0% |
| Celeb-reID-light | 72.0% | +20.0% (vs DeSKPro) |
| LaST | 79.0% | Significant margin |
| LTCC | SOTA | Competitive performance |
Ablation studies confirmed that a 100% masking ratio for cloth-related attributes yielded the best results, proving that clothing features act as noise in long-term re-identification tasks.
Conclusion
MADE demonstrates that masked attribute descriptions provide a powerful and efficient alternative to complex geometric modeling for cloth-changing Re-ID. By leveraging the flexibility of binary attribute vectors and the representational power of Vision Transformers, the framework sets a new standard for identifying individuals in dynamic, real-world environments. Future work may involve integrating Large Language Models (LLMs) to generate more nuanced descriptions and exploring cross-modality applications.
Summary updated