Harmonious convergence for confidence estimation in depth estimation and completion
Keywords: confidence estimation, monocular depth estimation, depth completion
TL;DR: We propose a harmonious convergence estimation approach for confidence estimation in the regression tasks, taking training consistency information into consideration.
Abstract: Confidence estimation for monocular depth estimation and completion is important for their deployment in real-world applications. Recent models for confidence estimation in these regression tasks mainly rely on the statistical characteristics of training and test data, while ignoring the information from the model training. We propose a harmonious convergence estimation approach for confidence estimation in the regression tasks, taking training consistency into consideration. Specifically, we propose an intra-batch convergence estimation algorithm with two sub-iterations to compute the training consistency for confidence estimation. A harmonious convergence loss is newly designed to encourage the consistency between confidence measure and depth prediction. Our experimental results on the NYU2 and KITTI datasets show improvements ranging from 10.91\% to 43.90\% across different settings in monocular depth estimation, and from 27.91\% to 45.24\% in depth completion, measured by Pearson correlation coefficients, justifying the effectiveness of the proposed method. We will release all the codes upon the publication of our paper.
Primary Area: applications to computer vision, audio, language, and other modalities
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 972
Loading