EUVIMEDEUVIMED entdecken
Crossref

Uncertainty-Aware Monocular Depth Estimation for Indoor Scenes Using a Hybrid Bayesian–CNN Framework

Divya Kiran, Ugra Mohan Roy

Journal of Innovative Image Processing · 12.05.2026 · 8 (2) · 644-664

DOI: 10.36548/jiip.2026.2.011 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Journal of Innovative Image Processing, online erschienen am 12.05.2026. Verfasst von Divya Kiran und 1 weiteren Autor:innen. Die Arbeit nennt 27 Literaturangaben.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Single RGB image monocular depth estimation is an inherently ill-posed task because three-dimensional data is lost when building an image. Despite the recent remarkable performance of deep learning-based models, most existing techniques are based on deterministic models that are weak in considering uncertainty, which restricts their application in complex indoor settings. To overcome this shortcoming, this paper introduces a hybrid probabilistic–deep learning architecture to estimate the depth of indoor monocular scenes, which combines Bayesian uncertainty representation with supervised convolutional neural network (CNN) refinement. As part of the proposed solution, a Bayesian Network is used to model the probabilistic dependence between image features and depth, generating pixel-wise posterior depth distributions that explicitly reflect estimation uncertainty. The following probabilistic depth priors are then used to train a supervised encoder-decoder CNN with an uncertainty-aware loss formulation that allows for smoother predictions of metric depth while retaining uncertainty information. This two-stage approach makes it less sensitive to ambiguous visual information and enhances depth estimation stability at a lower computational cost. Our proposed approach delivers an AbsRel of 0.080, RMSE of 0.290 and RMSE-log of 0.044 on the NYU Depth V2 dataset, along with a threshold accuracy of 93.0% with δ < 1.25. This shows comparable results with enhanced stability and uncertainty prediction for indoor depth estimation.

Zitieren

APA

Kiran, D., & Roy, U. M. (2026). Uncertainty-Aware Monocular Depth Estimation for Indoor Scenes Using a Hybrid Bayesian–CNN Framework. Journal of Innovative Image Processing, 8(2), 644-664. https://doi.org/10.36548/jiip.2026.2.011

Vancouver

Kiran D, Roy UM. Uncertainty-Aware Monocular Depth Estimation for Indoor Scenes Using a Hybrid Bayesian–CNN Framework. Journal of Innovative Image Processing. 2026;8(2):644-664. doi:10.36548/jiip.2026.2.011

BibTeX
@article{Kiran2026,
  author = {Kiran, Divya and Roy, Ugra Mohan},
  title = {Uncertainty-Aware Monocular Depth Estimation for Indoor Scenes Using a Hybrid Bayesian–CNN Framework},
  journal = {Journal of Innovative Image Processing},
  year = {2026},
  volume = {8},
  number = {2},
  pages = {644--664},
  doi = {10.36548/jiip.2026.2.011}
}
Kostenlos starten →