Crafting Monocular Cues and Velocity Guidance for Self-Supervised Multi-Frame Depth Learning

Self-supervised monocular methods can efficiently learn depth information of weakly textured surfaces or reflective objects. However, the accuracy is limited due to inherent ambiguity in geometric modeling. In contrast, multi-frame estimation improve thanks success Multi-View Stereo (MVS), which directly makes use constraints. Unfortunately, MVS often suffers from texture-less regions, non-Lambertian surfaces, and moving objects, especially real-world video sequences without known camera motion supervision. Therefore, we propose MOVEDepth, exploits MOnocular cues VElocity guidance Depth learning. Unlike existing that enforce consistency between depth, MOVEDepth boosts learning by addressing problems MVS. The key our approach utilize as a priority construct cost volume, adjust candidates volume under predicted velocity. We further fuse uncertainty results robust against multi-view geometry. Extensive experiments show achieves state-of-the-art performance: Compared with Monodepth2 PackNet, method relatively improves 20% 19.8% on KITTI benchmark. also generalizes more challenging DDAD benchmark, outperforming ManyDepth 7.2%. code available at https://github.com/JeffWang987/MOVEDepth.