Illumination Insensitive Monocular Depth Estimation Based on Scene Object Attention and Depth Map Fusion

Monocular depth estimation (MDE) is a crucial but challenging computer vision (CV) task which suffers from lighting sensitivity, blurring of neighboring depth edges, and object omissions. To address these problems, we propose an illumination insensitive monocular depth estimation method based on scene object attention and depth map fusion. Firstly, we design a low-light image selection algorithm, incorporated with the EnlightenGAN model, to improve the image quality of the training dataset and reduce the influence of lighting on depth estimation. Secondly, we develop a scene object attention mechanism (SOAM) to address the issue of incomplete depth information in natural scenes. Thirdly, we design a weighted depth map fusion (WDMF) module to fuse depth maps with various visual granularity and depth information, effectively resolving the problem of blurred depth map edges. Extensive experiments on the KITTI dataset demonstrate that our method effectively reduces the sensitivity of the depth estimation model to light and yields depth maps with more complete scene object contours.