EUVIMEDEUVIMED entdecken
Crossref

Probabilistic Monocular and Multiview Network With Reliability-Aware Cost Volume Fusion for Depth Estimation

Anjie Wang, Guibiao Liao, Zhongyuan Zhao, Ruotong Li, Zhijun Fang, Siwei Ma, Jenq-Neng Hwang

IEEE Transactions on Instrumentation and Measurement · 2026 · 75 · 5014812-5014812

DOI: 10.1109/tim.2026.3711298 ↗

Kurz gefasst

Fachartikel in der Zeitschrift IEEE Transactions on Instrumentation and Measurement, 2026. Verfasst von Anjie Wang und 6 weiteren Autor:innen. Die Arbeit nennt 54 Literaturangaben.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Recent advances in monocular depth estimation suggest that single-view priors and multiview geometry are complementary: multiview matching yields high spatial accuracy but often breaks on textureless/reflective regions and dynamic objects, whereas single-view networks provide strong semantic regularization yet may suffer from geometric inconsistency. To bridge this gap, we propose a probabilistic monocular and multiview guided network (ProMaGNet), a self-supervised framework for robust multiframe monocular depth estimation that unifies probabilistic hypothesis generation with reliability-aware cost aggregation and fusion. ProMaGNet predicts a per-pixel Gaussian prior in inverse depth from the reference frame, enabling efficient prior-driven sampling to construct compact depth hypotheses and cost volumes. We build a raw multiview cost volume via feature warping and matching, and a monocular unary cost volume as the negative log-likelihood under the predicted prior. To explicitly capture cross-view disagreement, we further compute a multiview consistency cue using the variance across source views. A lightweight 3-D reliability network jointly predicts candidate reliability and a pixel–candidate fusion gate, enabling our reliability-aware cost volume modulation (RCVM) to adaptively balance multiview evidence and monocular priors and suppress unreliable hypotheses in occluded or dynamic regions. Moreover, an uncertainty-reweighted photometric objective mitigates misleading supervision from ambiguous areas and stabilizes training. Experiments on KITTI and Cityscapes demonstrate that ProMaGNet consistently outperforms state-of-the-art self-supervised baselines, with notable improvements in both static and dynamic scenes.

Zitieren

APA

Wang, A., Liao, G., Zhao, Z., Li, R., Fang, Z., Ma, S., & Hwang, J. (2026). Probabilistic Monocular and Multiview Network With Reliability-Aware Cost Volume Fusion for Depth Estimation. IEEE Transactions on Instrumentation and Measurement, 75, 5014812-5014812. https://doi.org/10.1109/tim.2026.3711298

Vancouver

Wang A, Liao G, Zhao Z, Li R, Fang Z, Ma S, et al. Probabilistic Monocular and Multiview Network With Reliability-Aware Cost Volume Fusion for Depth Estimation. IEEE Transactions on Instrumentation and Measurement. 2026;75:5014812-5014812. doi:10.1109/tim.2026.3711298

BibTeX
@article{Wang2026,
  author = {Wang, Anjie and Liao, Guibiao and Zhao, Zhongyuan and Li, Ruotong and Fang, Zhijun and Ma, Siwei and Hwang, Jenq-Neng},
  title = {Probabilistic Monocular and Multiview Network With Reliability-Aware Cost Volume Fusion for Depth Estimation},
  journal = {IEEE Transactions on Instrumentation and Measurement},
  year = {2026},
  volume = {75},
  pages = {5014812--5014812},
  doi = {10.1109/tim.2026.3711298}
}
Kostenlos starten →