Weakly supervised free-space segmentation by fusing spatial priors and region features for auto-driving

Weakly supervised semantic segmentation can significantly alleviate the annotation burden of the pixel-level collection used in full supervision. However, most existing works are based on simple images and only include a few tags, which are not applicable to free-space segmentation under complex driving scenes. In this study, we develop an effective weakly supervised framework with image-level label for free-space segmentation by incorporating the spatial priors with geometric context of road. The proposed method jointly leverages the locating capability of image tags and low-level structure information of superpixels, and fuses spatial weights and superpixel region features by a region-based clustering method. These strategies guarantee the satisfying segmentation results while facing the complex driving scenes. Extensive experimental results on the Cityscapes datasets show that our work outperforms some competing weakly supervised methods in terms of intersection over union (IoU), and is a feasible way to perform free-space segmentation in complex environments for auto-driving application.