EUVIMEDAtrasti EUVIMED
Open access · CC BY 4.0

DCE‐MVSNet: Unsupervised Multi‐View Stereo With a Learnable Depth Confidence Evaluator via Adaptive Volume Rendering

Liangliang Li, Guihua Liu, Feng Xu, Wenjin Liao

IET Image Processing · 2026-01 · 20 (1)

DOI: 10.1049/ipr2.70520 ↗

At a glance

Journal article in IET Image Processing, published online on 2026-11-14. Written by Liangliang Li and 3 co-authors. 51 references. The full text is freely available (open access, CC BY 4.0). Funded by: Science and Technology Department of Sichuan Province.

SantraukaBibliografiniai duomenysCiteRelated studies

Abstract (original language)

In unsupervised multi‐view stereo (MVS), accurately evaluating the reliability of predicted depth remains a core challenge. Existing methods rely on the photometric consistency assumption to construct loss functions, making it difficult to handle unreliable supervision signals caused by weak textures, repetitive textures, specular reflections and occlusions. Although RC‐MVSNet pioneered the introduction of neural rendering into unsupervised MVS, using rendered depth as pseudo ground truth to provide additional supervision, it essentially fails to solve the problem of how to accurately evaluate the reliability of predicted depth. To this end, we propose DCE‐MVSNet, which redefines the neural rendering network as a learnable depth confidence evaluator to provide reliable training targets for the MVS network. Specifically, we treat the depth hypothesis range used by the MVS network for predicting depth maps as the volume sampling range for neural rendering. Based on the concentration of the probability distribution corresponding to the depth hypotheses, we adaptively employ either a precise sampling strategy based on the cumulative distribution function or a hybrid exploration sampling strategy. Through this approach, the neural rendering network can identify regions where the photometric consistency loss is reliable and guide the MVS network to correct its predictions via confidence‐weighted rendering errors. Extensive experiments demonstrate that DCE‐MVSNet outperforms the current state‐of‐the‐art neural rendering‐based unsupervised method, RC‐MVSNet, improving accuracy, completeness and overall metrics by 2.27%, 6.10% and 3.77%, respectively, on the DTU dataset, achieving the best completeness among end‐to‐end unsupervised MVS methods. Furthermore, it achieves the second‐highest mean F‐score among the compared end‐to‐end unsupervised MVS methods on both the intermediate and advanced subsets of Tanks & Temples.

Funding

  • Science and Technology Department of Sichuan Province

Cite

APA

Li, L., Liu, G., Xu, F., & Liao, W. (2026). DCE‐MVSNet: Unsupervised Multi‐View Stereo With a Learnable Depth Confidence Evaluator via Adaptive Volume Rendering. IET Image Processing, 20(1). https://doi.org/10.1049/ipr2.70520

Vancouver

Li L, Liu G, Xu F, Liao W. DCE‐MVSNet: Unsupervised Multi‐View Stereo With a Learnable Depth Confidence Evaluator via Adaptive Volume Rendering. IET Image Processing. 2026;20(1). doi:10.1049/ipr2.70520

BibTeX
@article{Li2026,
  author = {Li, Liangliang and Liu, Guihua and Xu, Feng and Liao, Wenjin},
  title = {DCE‐MVSNet: Unsupervised Multi‐View Stereo With a Learnable Depth Confidence Evaluator via Adaptive Volume Rendering},
  journal = {IET Image Processing},
  year = {2026},
  volume = {20},
  number = {1},
  doi = {10.1049/ipr2.70520}
}
Start for free →