MLDSE: Multi-Task Learning for Depth and Segmentation Estimation

Monocular depth estimation remains a challenging task due to the susceptibility to photometric errors and inaccuracies in the depth boundaries. Recent methods use a novel multi task framework that combines depth estimation with the segmentation task to leverage the inherent relationship between depth and object boundaries, refining depth estimates and enforcing depth discontinuities. In this paper, we investigate three novel attention modules that operate on both spatial and channel dimensions and propose a Cross-Task Feature Propagation Unit that effectively enables enhanced information transfer across tasks. Furthermore, we conduct an in-depth analysis of the multi-scale loss of segmentation and its influence on depth estimation accuracy. Experiments performed on the KITTI datasets confirm the efficacy of our framework, producing results that are competitive with those obtained using state-of-the-art techniques.