Image Purification Networks: Real-time Style Transfer with Semantics through Feed-forward Synthesis

摘要

Image synthesis has been widely accepted as a cost effective way to learn models because it provides training sets that are large, diverse and accurately labeled. However, the realism of the synthetic image is not enough, this affects generalization on naturalistic test image. In an attempt to address this issue, previous methods learn a model to improve the realism of synthetic image. Differently, from previous methods, we take the first step towards purifying the naturalistic image to weaken the influence of light and convert the distribution of an outdoor naturalistic image through a real-time style transfer task to that of indoor synthetic image. This paper proposes, therefore a real-time image purification networks that transfer style information with semantics through a feed-forward synthesis. Results from our experiments demonstrate that images purified through the proposed networks architecture trained models for gaze estimation more accurately on cross-datasets over using raw naturalistic images and when compared to baseline methods.