On Monocular Depth Estimation and Uncertainty Quantification Using Classification Approaches for Regression

Abstract

Monocular depth is important in many tasks, such as 3D reconstruction and autonomous driving. Deep learning based models achieve state-of-the-art performance in this field. A set of novel approaches for estimating monocular depth consists of transforming the regression task into a classification one. However, there is a lack of detailed descriptions and comparisons for Classification Approaches for Regression (CAR) in the community and no in-depth exploration of their potential for uncertainty estimation. To this end, this paper will introduce a taxonomy and summary of CAR approaches, a new uncertainty estimation solution for CAR, and a set of experiments on depth accuracy and uncertainty quantification for CAR-based models on KITTI dataset. The experiments reflect the differences in the portability of various CAR methods on two backbones. Meanwhile, the newly proposed method for uncertainty estimation can outperform the ensembling method with only one forward propagation.

Discretization-Induced Dirichlet Posterior for Robust Uncertainty Quantification on Regression

“…The input of AuxUE can be the input, output, or intermediate features of f ω , we here simplify it to the image x (i) for brevity. and σ Θ2 are based on the basic AuxUEs such as Confid-Net (Corbière et al 2019), BayesCap (Upadhyay et al 2022) and SLURP (Yu, Franchi, and Aldea 2021) depending on the tasks. The input of AuxUE can be the input, output, or intermediate features of f ω and it depends on the design of the basic AuxUEs, which is not the focus of this paper.…”

“…Discretization on prediction errors To mitigate numerical bias due to imbalanced data in our prediction error estimation, we employ a balanced discretization approach. Discretization is widely applied in classification approaches for regression (Yu, Franchi, and Aldea 2022). The popular discretization methods can be generally divided into hand-…”

“…As one of the most effective methods, Auxiliary Uncertainty Estimators (AuxUE) (Corbière et al 2019;Yu, Franchi, and Aldea 2021;Jain et al 2021;Corbière et al 2021;Besnier et al 2021;Upadhyay et al 2022;Shen et al 2023) aim to obtain uncertainty estimates without affecting the main task performance. AuxUEs are DNNs that rely on the main task models used for estimating the uncertainty of the main task prediction.…”

“…For vision regression tasks, basic AuxUE addresses only aleatoric uncertainty estimation (Yu, Franchi, and Aldea 2021). Recent works (Upadhyay et al 2022;Qu et al 2022) aim to improve the generalization ability of the basic Aux-UEs.…”

“…To further explore robust aleatoric and epistemic uncertainty estimation in vision regression tasks, in this work, we propose a novel uncertainty quantification solution based on AuxUE. For estimating aleatoric uncertainty, we follow the approach of previous works such as (Nix and Weigend 1994;Kendall and Gal 2017;Yu, Franchi, and Aldea 2021;Upadhyay et al 2022) and model the heteroscedastic noise using different distribution assumptions. For epistemic uncertainty quantification, we apply a discretization approach to the continuous prediction errors of the main task.…”