Improving Fashion Landmark Detection by Dual Attention Feature Enhancement

Abstract

Fashion landmark detection is a fundamental problem in visual fashion analyze, which aims at locating the precise coordinates of functional key points defined on clothes. Dozens of deep learning-based methods are proposed to address this problem. How to extract adequate and effective features is a critical point for this challenging task. In this paper, we propose the Dual Attention Feature Enhancement(DAFE) module, which strengthens the extracted features by adaptively reusing low-level image details and emphasizing informative parts. First, DAFE enhances the pixel-wise information through capturing the spatial details from low-level features by the guidance of attention matrix, which is generated from high-level ones. Second, DAFE emphasizes task-related features by modeling long-range relationships between channels. Experimental experiments on Deepfashion and FLD datasets demonstrate that our method achieves state-of-the-art performance, and our approach also achieves competitive results on Deepfashion2 Landmark Estimation Challenge.