Semantic Image Segmentation on Snow Driving Scenarios

摘要

The challenges of driving on snow and ice roads bring out a demand for object detection and drivable area segmentation in snowy environments. Semantic segmentation techniques have been able to achieve good results provided that the models are well-trained based on the appropriate dataset. However, no current driving dataset exists that contains adequate images in snowy environments. To address this issue, we introduce our snowy driving dataset to train and test models for pixel-wise semantic labeling. This snowy driving dataset consists of both real and synthetic samples with 11 classes. We conduct comparative experiments based on a series of the proposed dataset, as well as provide statistics and visual results to show improvement.