Weakly-Supervised Semantic Segmentation via Self-training

Weakly-supervised semantic segmentation with image tags is a challenging computer vision task. Unlike pixel-level masks, image tags give high level semantic information, without low level appearance information. In this paper, we propose an iteratively self-training framework to bridge this two information, which expand and refine the pseudo- labels with training process going. Initial masks are generated from classification network. In the top-down step, rendered images and its labels as well as spatially weight loss are added to jointly training the model for alleviate the effect of inaccurate object masks. Then in the bottom-up step, an adaptive threshold to the confidence model predictions to keep predicted masks reliable. The top-down and bottom-up steps are conducted iteratively to extract the fine object mask. Experiments on our self-build dataset and GTA5 to CityScapes demonstrate the effectiveness of proposed framework.