Robust Semantic Segmentation for Street Fashion Photos

In this paper, we aim to produce the state-of-the-art semantic segmentation for street fashion photos with three contributions. Firstly, we propose a high-performance semantic segmentation network that follows the encoder-decoder structure. Secondly, we propose a guided training process using multiple auxiliary losses. And thirdly, the 2D max-pooling-based scaling operation to produce segmentation feature maps for the aforementioned guided training process. We also propose mIoU+ metric taking noise into account for better evaluation. Evaluations with the ModaNet data set show that the proposed network achieves high benchmark results with less computational cost compared to ever-proposed methods.