Deep garment image matting for a virtual try-on system

To improve online shopping experience, many fashion retailers try to provide high quality garment images, capturing fine details as well as various opacities. A skilled operator can deliver a satisfactory result using manual segmentation tools, but it is challenging to scale up this process to address seasonal demands. To balance the quality and the processing cost, we investigate the use of a deep learning based matting technique that can produce a high quality alpha map from an approximate garment segmentation. The proposed model adopts the deep image matting model, but we replace the refinement network with a sequence of recursive convolutional network (RCN) units. Our main motivation for this modification is that the fine garment details created by different materials are represented better with the mixture of the image features from different scales. Therefore, we need to construct deeper convolutional layers for better scale analysis but we also need to maintain the number of unknowns low as producing training data is expensive. The proposed RCN based refinement network can address these conflicting restrictions well and our experiments demonstrate that it can achieve a lower training loss and produce better prediction results than the baseline refinement model under the same training condition.