Abstract
Visual Attribute Prediction (VAP) is a fundamental task in fashion AI, yet it remains challenging due to label sparsity, high annotation costs, and domain dependency. This paper introduces the Multi-Teacher-Single-Student (MTSS) framework, a semi-supervised approach that trains specialized teacher networks for distinct attribute types and distills their collective knowledge into a single student model. By leveraging unlabeled data and employing label embedding techniques, MTSS achieves state-of-the-art performance on major benchmarks without requiring auxiliary domain-specific metadata.
Introduction
Visual Attribute Prediction involves identifying fine-grained characteristics such as color, pattern, and cut. Despite its importance in e-commerce, VAP faces three primary hurdles: label sparsity (the "long-tail" problem), the prohibitive expense of manual annotation, and a reliance on domain-specific "crutches" like clothing landmarks. This research proposes a domain-agnostic framework that utilizes unlabeled data to build a robust, accurate model capable of operating in real-world, noisy environments.
Methodology
The proposed MTSS framework follows a two-stage training process designed to maximize expertise and generalizability.
Stage 1: Training the Experts (Teachers)
The first stage involves training multiple "Teacher" networks, each serving as a domain expert for a specific attribute category (e.g., Color or Pattern). Instead of standard binary classification, the authors utilize Label Embedding (LE) to map images and labels into a shared latent space. This approach captures the conceptual hierarchy and relationships between attributes, such as the similarity between different patterns.
To further refine these experts, a modified focal loss is applied to metric learning, forcing the teachers to focus on "hard" examples that are frequently misclassified.
Stage 2: The Unified Student
In the second stage, a single "Student" network is trained to mimic the feature distributions of all specialized teachers. This is achieved through Knowledge Distillation, where the student attempts to match the embeddings produced by the experts.
A key advantage of this stage is the ability to incorporate vast amounts of unlabeled data, as the student only requires the teachers' outputs for training. To maintain quality, an outlier suppression mechanism (controlled by hyperparameter $\beta$) prevents the student from learning from uncertain or noisy teacher predictions.
Experimental Setup
The MTSS model was evaluated across four diverse datasets:
- DeepFashion: The standard benchmark for fashion attribute prediction.
- iMaterialist-Fashion-2018: A large-scale dataset from the Kaggle competition.
- DARN: Utilized for cross-domain retrieval testing.
- FiccY: A new private dataset containing 520,000 high-resolution, real-world e-commerce images, providing a testbed for model robustness against user-uploaded "noisy" content.
Discussion and Key Insights
The research yielded several critical insights into attribute prediction:
- Embedding vs. Classification: Label Embedding consistently outperformed Binary Cross-Entropy (BCE), particularly in data-constrained scenarios, by capturing the inherent structure of attribute classes.
- Teacher-Student Dynamics: While the student model often surpasses individual teachers due to the regularizing effect of distillation and additional data, its performance is ultimately bounded by the teachers' initial quality.
- Robustness: Integrating "noisy" unlabeled data during the distillation phase significantly improved the model's resistance to image corruptions (e.g., blur) without compromising performance on clean benchmark data.
Results
The MTSS framework demonstrated superior performance on the DeepFashion benchmark, achieving a Top-3 Recall of 58.02%. This result outperforms previous methods that lacked auxiliary labels and remains competitive with those that utilize them. Furthermore, the model showed significant gains in cross-domain adaptation and noise resistance, with a 2.33% accuracy improvement on corrupted images when trained with additional unlabeled data.
Conclusion
The MTSS framework provides a scalable, domain-agnostic solution for visual attribute prediction. By effectively combining specialized expertise with semi-supervised distillation, it addresses the long-tail problem and reduces reliance on expensive annotations. Future work will focus on optimizing teacher quality and exploring advanced sampling strategies for metric learning to further enhance performance on rare attributes.
Summary updated