Bayesian Deep Neural Networks for Supervised Learning of Single-View Depth

TL;DR

This paper evaluates scalable uncertainty quantification methods, like MC dropout and deep ensembles, for single-view depth learning, highlighting the effectiveness of dropout in all encoder layers for robotic perception.

Contribution

It demonstrates that applying dropout across all encoder layers improves depth uncertainty estimation, matching deep ensembles with less memory, and explores its use in two-view motion estimation.

Findings

Dropout in all encoder layers yields better depth uncertainty estimates.

Dropout-based approach performs similarly to deep ensembles with lower memory.

Depth uncertainty improves two-view relative motion estimation with real scale.

Abstract

Uncertainty quantification is essential for robotic perception, as overconfident or point estimators can lead to collisions and damages to the environment and the robot. In this paper, we evaluate scalable approaches to uncertainty quantification in single-view supervised depth learning, specifically MC dropout and deep ensembles. For MC dropout, in particular, we explore the effect of the dropout at different levels in the architecture. We show that adding dropout in all layers of the encoder brings better results than other variations found in the literature. This configuration performs similarly to deep ensembles with a much lower memory footprint, which is relevant forapplications. Finally, we explore the use of depth uncertainty for pseudo-RGBD ICP and demonstrate its potential to estimate accurate two-view relative motion with the real scale.