Triple Attention Network for Clothing Parsing

Clothing parsing has been actively studied in the vision community in recent years. Inspired by the color coherence for clothing and the self-attention mechanism, this paper proposes a Triple Attention Network (TANet) equipped with a color attention module, a position attention module and a channel attention module, to facilitate fine-grained segmentation of clothing images. Concretely, the color attention module is introduced for harvesting color coherence, which selectively aggregates the color feature of clothing. The position attention module and the channel attention module are designed to emphasize the semantic interdependencies in spatial and channel dimensions respectively. The outputs of the three attention modules are incorporated to further improve feature representation which contributes to more precise clothing parsing results. The proposed TANet has achieved 69.54\(\%\) mIoU - a promising clothing parsing performance on ModaNet, the latest large-scale clothing parsing dataset. Especially, the color attention module is also demonstrated to bring semantic consistency and precision improvement obviously. The source code is made available in the public domain.