Implementation of Off-Road Panoptic-Segmentation Under Dynamic Lighting Conditions

Understanding off-road environments, especially forests, is a challenging problem in computer vision due to factors including illumination, features extraction and ambiguity. This work aims at analysing the off-road environments by implementing Panoptic segmentation (PS) architecture taking into consideration the dynamic lighting conditions. To assist this, a novel off-road Panoptic dataset is also presented and consists of 285 annotated forestry images in varying illumination conditions. To handle the problem of illumination variations, Shadow-removal/exposure correction techniques are further experimented with. We include these correction techniques as a preprocessing step not only to improve the quality of feature extraction but also to reduce the ambiguity caused by varying lighting conditions. The experiments are further supported with qualitative and quantitative analysis. The source code and dataset is available at github.