Uncertainty Diffusion: Parameter-Efficient Depth Refinement via Uncertainty-Guided Diffusion Models


Jeng-Huo Tzeng (National Yang Ming Chiao Tung University), Chuan-Yuan Huang (National Yang Ming Chiao Tung University), Kuan-Wen Chen (National Yang Ming Chiao Tung University)
The 35th British Machine Vision Conference
PDFPosterVideo (Right click to download)

Abstract

We present Uncertainty Diffusion, a model-agnostic framework for refining monocular depth maps by integrating pixel-wise uncertainty into a diffusion-based process. Our method adaptively focuses refinement on regions with low prediction confidence by leveraging an uncertainty-guided sampling mechanism, enabling targeted and effective depth enhancement. For domain adaptation, only a lightweight refinement network is fine-tuned while keeping the base model fixed, resulting in a parameter-efficient adaptation strategy. Extensive experiments on NYU Depth V2, DIODE, and SUN RGB-D demonstrate consistent improvements across diverse baseline models, including recent large-scale approaches. Notably, our framework achieves up to a 61\% reduction in log10 error for domain adaptation, all without retraining the base models. These results highlight the practicality and versatility of Uncertainty Diffusion for robust monocular depth estimation in varied environments.

Citation

@inproceedings{Tzeng_2025_BMVC,
author    = {Jeng-Huo Tzeng and Chuan-Yuan Huang and Kuan-Wen Chen},
title     = {Uncertainty Diffusion: Parameter-Efficient Depth Refinement via Uncertainty-Guided Diffusion Models},
booktitle = {36th British Machine Vision Conference 2025, {BMVC} 2025, Sheffield, UK, November 24-27, 2025},
publisher = {BMVA},
year      = {2025},
url       = {https://bmva-archive.org.uk/bmvc/2025/assets/papers/Paper_128/paper.pdf}
}


Copyright © 2025 The British Machine Vision Association and Society for Pattern Recognition
The British Machine Vision Conference is organised by The British Machine Vision Association and Society for Pattern Recognition. The Association is a Company limited by guarantee, No.2543446, and a non-profit-making body, registered in England and Wales as Charity No.1002307 (Registered Office: Dept. of Computer Science, Durham University, South Road, Durham, DH1 3LE, UK).

Imprint | Data Protection