Depth estimation is an image translation problem that predicts depth maps for a given camera image and has fostered research in various applications including self-driving vehicles. Self-supervised depth estimation methods are of particular interest since ground truth LIDAR depth is expensive to acquire and instead use view synthesis as weaker supervision. Generally, the produced depth maps to date are only point estimates of an underlying depth distribution due to randomness in model training, resulting in noisy depth estimates that can propagate errors and lead to inaccurate or fatal decisions in real-world applications. Recent interest has been sparked in reducing such noise by modeling the uncertainty of depth estimates. Empirical uncertainty strategies seek to predict uncertainty via statistical methods by treating independent models as black box predictors. Of particular interest are predictive strategies that seek to learn the inherent uncertainty of a depth model. For example, student-teacher frameworks train one network to learn the depth output distribution of another. Such methods are desirable due to the advantage of requiring fewer training and space resources compared to other empirical methods. In this work, we study self-supervised depth models with a U-Net architecture that outputs depths at multiple scales. In particular, we explore a novel predictive uncertainty model that only has access to these scales and the U-Net bottleneck feature. We evaluate and discuss the novel method alongside other uncertainty strategies on the KITTI dataset.

50 pages

Thesis Committee:
Jean Oh (Chair)
Ji Zhang

Srinivasan Seshan, Head, Computer Science Department
Martial Hebert, Dean, School of Computer Science

Return to: SCS Technical Report Collection
School of Computer Science

This page maintained by [email protected]