Semi-Supervised Disparity Estimation with Deep Feature Reconstruction

TL;DR

This paper introduces a semi-supervised approach combining supervised and self-supervised learning to improve disparity estimation across domains, utilizing deep feature reconstruction to address photometric loss limitations.

Contribution

It presents a novel semi-supervised pipeline that adapts DispNet to real-world data and explores deep feature reconstruction as an effective supervisory signal.

Findings

Improved domain generalization in disparity estimation

Deep feature reconstruction enhances supervision beyond photometric loss

Effective adaptation of DispNet to real-world data

Abstract

Despite the success of deep learning in disparity estimation, the domain generalization gap remains an issue. We propose a semi-supervised pipeline that successfully adapts DispNet to a real-world domain by joint supervised training on labeled synthetic data and self-supervised training on unlabeled real data. Furthermore, accounting for the limitations of the widely-used photometric loss, we analyze the impact of deep feature reconstruction as a promising supervisory signal for disparity estimation.