Models for unsupervised monocular depth estimation (MDE) have gained much attention due to recent breakthroughs and the ability to train with unlabeled data. Despite the state-of-the-art methods performing well on depth prediction benchmarks, certain artifacts and their performance compared to their supervised counterparts make them less favorable in certain domains. This thesis analyzes these models and presents a set of methods for improvement which can be applied in the training process.Recent papers in unsupervised MDE focus on increasing performance metrics on the KITTI benchmark. We show that the results from these methods can be further improved by (i) providing synthetic training data via the game engine Grand Theft Auto V (GTAV) and (ii) applying data augmentation techniques that are consistent with the camera intrinsic parameters of the model.
【 预 览 】
附件列表
Files
Size
Format
View
Unsupervised monocular depth estimation: Learning to generalize