DepSAM: Monocular Depth Estimation with SAM in Night-time Scenes

This paper tackles photometric inconsistency in night-time monocular depth estimation, focusing on underexposure in unlit areas and overexposure from materials or proximity to lights, causing depth estimation inaccuracies. DepSAM: Depth estimation network leveraging the Segment Anything Model (SAM) to propose three enhancements: 1) leveraging SAM to create semantic masks for depth consistency in dark regions, 2) designing an AICE module using a light mask learned by neural networks to handle exposure issues, and 3) introducing a semantic-depth inflation loss and relative depth consistency constraint loss that utilize semantic boundaries and vehicle size data to refine depth ratios and totals within scenes. We have achieved state-of-the-art performance on the benchmark tasks, with enhanced preservation of object integrity and clearer definition of object outlines. Our codes will be available soon.