DiffSF: Diffusion Models for Scene Flow Estimation

YushanZhang, Bastian Wandt, Maria Magnusson Michael Felsberg
Computer Vision Laboratory, Linköping University   
Paper Code
Source Image
Method Overview

TLDR

Aiming at improving accuracy while additionally providing an estimate for uncertainty, we propose DiffSF that combines transformer-based scene flow estimation with denoising diffusion models.

Method

We propose Diffusion Models for Scene Flow Estimation (DiffSF). Our approach consists of three main contributions:

  1. We introduce DiffSF, leveraging diffusion models to solve the full scene flow estimation problem, where the inherent noisy property of the diffusion process filters out noisy data, thus, increasing the focus on learning the relevant patterns.
  2. DiffSF introduces randomness to the scene flow estimation task, which allows us to predict the uncertainty of the estimates without being explicitly trained for this purpose.
  3. We develop a novel architecture that combines transformers and diffusion models for the task of scene flow estimation, improving both accuracy and robustness for a variety of datasets.

BibTeX


      @article{zhang2024diffsf,
        title={DiffSF: Diffusion Models for Scene Flow Estimation},
        author={Zhang, Yushan and Wandt, Bastian and Magnusson, Maria and Felsberg, Michael},
        journal={arXiv preprint arXiv:2403.05327},
        year={2024}
      }