FashionLens: Toward Versatile Fashion Image Retrieval via Task-Adaptive Learning

Researchers introduced FashionLens, a unified fashion image retrieval framework built for multiple real-world query types.

It is backed by U-FIRE, a new benchmark that consolidates fashion datasets and adds fresh test sets for broader evaluation.

FashionLens uses task-adaptive learning, including query calibration and reweighting, to handle data imbalance and diverse retrieval goals.

The system reports state-of-the-art results and releases code and data, aiming to improve practical fashion search and generalization.