Get started
Work

Hypergraph-Based Depth Map Quality Assessment Using Monocular Depth Estimation

Year: 2026
Type: article
Abstract: As depth data is increasingly utilized in wide-ranging applications, the importance of depth map quality assessment (DQA) has grown significantly. However, compared to the image quality assessment (IQA), the methods for DQA still require more exploration. Due to the modality gap between depth maps and images, traditional IQA metrics lack consistency and are unsuitable for DQA tasks. Additionally, the absence of ground truth for depth quality scores makes conducting end-to-end training for DQA methods challenging. To tackle these challenges, we propose an intuition to evaluate the performance of DQA methods through monocular depth estimation (MDE). Specifically, we design a hypergraph-based color-reference DQA method (HyperDQA), which effectively models the cross-modality correlations between the reference images and depth maps and generates detailed depth quality score maps (QSM) through hypergraph computation. Moreover, we present an end-to-end two-stage framework to evaluate the DQA methods, which can explicitly correct erroneous pixel-level depth to obtain a more accurate final depth map during the “assessment-fusion” stage. The experimental results and comparisons indicate that MDE metrics show better consistency than IQA metrics on the DQA task. Moreover, our proposed HyperDQA-based “assessment-fusion” training paradigm and QSM loss can significantly improve the depth map quality and outperform existing state-of-the-art MDE methods on publicly used datasets. The extensive experiments testing on zero-shot datasets demonstrate the superiority of the generalization ability of our proposed fusion strategy.
DOI: 10.1109/tmm.2026.3726244
Institution: Tsinghua University
Language: English
Open Access status: hybrid
Estimated APC paid: $2,645
OpenAlex ID: W7203909083
FWCI:
Cites:
Cited by:
Related to: 10