Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Stereo-Matching Knowledge Distilled Monocular Depth Estimation Filtered by Multiple Disparity Consistency

Woonghyun Ka, Jae Young Lee, Jaehyun Choi, Junmo Kim

TL;DR

Self-supervised monocular depth estimation often suffers from errors in pseudo-depth generated by stereo-matching networks. This work proposes a GT-free filtering mechanism based on consistency across multiple disparity maps obtained via disparity plane sweep, producing a weight map to down-weight unreliable regions during training. The weight map modulates the depth regression loss, enabling the monocular network to learn from accurate pseudo-depth without additional GT or stereo-confidence training. Experiments on KITTI Eigen split and Cityscapes demonstrate improved accuracy and robustness across backbone and stereo-network configurations, with qualitative gains at object boundaries and in challenging regions.

Abstract

In stereo-matching knowledge distillation methods of the self-supervised monocular depth estimation, the stereo-matching network's knowledge is distilled into a monocular depth network through pseudo-depth maps. In these methods, the learning-based stereo-confidence network is generally utilized to identify errors in the pseudo-depth maps to prevent transferring the errors. However, the learning-based stereo-confidence networks should be trained with ground truth (GT), which is not feasible in a self-supervised setting. In this paper, we propose a method to identify and filter errors in the pseudo-depth map using multiple disparity maps by checking their consistency without the need for GT and a training process. Experimental results show that the proposed method outperforms the previous methods and works well on various configurations by filtering out erroneous areas where the stereo-matching is vulnerable, especially such as textureless regions, occlusion boundaries, and reflective surfaces.

Stereo-Matching Knowledge Distilled Monocular Depth Estimation Filtered by Multiple Disparity Consistency

TL;DR

Self-supervised monocular depth estimation often suffers from errors in pseudo-depth generated by stereo-matching networks. This work proposes a GT-free filtering mechanism based on consistency across multiple disparity maps obtained via disparity plane sweep, producing a weight map to down-weight unreliable regions during training. The weight map modulates the depth regression loss, enabling the monocular network to learn from accurate pseudo-depth without additional GT or stereo-confidence training. Experiments on KITTI Eigen split and Cityscapes demonstrate improved accuracy and robustness across backbone and stereo-network configurations, with qualitative gains at object boundaries and in challenging regions.

Abstract

In stereo-matching knowledge distillation methods of the self-supervised monocular depth estimation, the stereo-matching network's knowledge is distilled into a monocular depth network through pseudo-depth maps. In these methods, the learning-based stereo-confidence network is generally utilized to identify errors in the pseudo-depth maps to prevent transferring the errors. However, the learning-based stereo-confidence networks should be trained with ground truth (GT), which is not feasible in a self-supervised setting. In this paper, we propose a method to identify and filter errors in the pseudo-depth map using multiple disparity maps by checking their consistency without the need for GT and a training process. Experimental results show that the proposed method outperforms the previous methods and works well on various configurations by filtering out erroneous areas where the stereo-matching is vulnerable, especially such as textureless regions, occlusion boundaries, and reflective surfaces.
Paper Structure (11 sections, 3 equations, 4 figures, 3 tables)

This paper contains 11 sections, 3 equations, 4 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Proposed Method
  4. Preliminaries and Concept
  5. Overall Framework
  6. Generating Weight Map
  7. Loss Function
  8. Experiments
  9. Experimental Setup
  10. Results
  11. Conclusions

Figures (4)

  • Figure 1: Principle to obtain depth from a stereo camera system.
  • Figure 2: Overall framework of the proposed method.
  • Figure 3: Disparity profile observation.
  • Figure 4: Qualitative comparison with the previous methods on the KITTI Eigen split dataset. $d_{s}^{0}$ and $W$ denote the predicted disparity map of the stereo-matching network watson2020learning and the weight map of the proposed method (Ours), respectively.