Fashion Detection and Search via Decoupled Anchor-free Framework

Clothing image search serves as a pivotal technique for efficiently retrieving the most relevant clothing items based on customer queries. The prevailing approaches consist of two serial steps, which first detect and crop out the target clothing, and then feed into a network for similarity learning. Alternatively, there are also one-step approaches that incorporate clothing detection and search within an end-to-end framework. But this approach may encounter optimization contradictions during the training phase, given the disparate optimization requirements of the clothing detection and Re-ID branches. In this work, we propose an end-to-end network for joint clothing detection and search by adding a Re-ID branch based on the attention mechanism in parallel. At the same time, in order to alleviate the conflict between the two branches, we adopt a global context decoupling module to effectively separate the deep features obtained by the backbone network into different branch-specific representations. Extensive results verify the superiority of our proposed method.