Many classical and learning-based optical flow methods rely on hierarchical concepts to improve both accuracy and robustness. However, one of the currently most successful approaches -- RAFT -- hardly exploits such concepts. In this work, we show that multi-scale ideas are still valuable. More precisely, using RAFT as a baseline, we propose a novel multi-scale neural network that combines several hierarchical concepts within a single estimation framework. These concepts include (i) a partially shared coarse-to-fine architecture, (ii) multi-scale features, (iii) a hierarchical cost volume and (iv) a multi-scale multi-iteration loss. Experiments on MPI Sintel and KITTI clearly demonstrate the benefits of our approach. They show not only substantial improvements compared to RAFT, but also state-of-the-art results -- in particular in non-occluded regions.
%0 Conference Paper
%1 Jahedi2022_ICIP
%A Jahedi, Azin
%A Mehl, Lukas
%A Rivinius, Marc
%A Bruhn, Andrés
%B IEEE International Conference on Image Processing (ICIP)
%D 2022
%K 2022 B04 sfbtrr161 visus visus:bruhnas visus:mehlls
%R 10.48550/arXiv.2207.12163
%T Multi-Scale RAFT: combining hierarchical concepts for learning-based optical flow estimation
%U /brokenurl# https://doi.org/10.48550/arXiv.2207.12163
%X Many classical and learning-based optical flow methods rely on hierarchical concepts to improve both accuracy and robustness. However, one of the currently most successful approaches -- RAFT -- hardly exploits such concepts. In this work, we show that multi-scale ideas are still valuable. More precisely, using RAFT as a baseline, we propose a novel multi-scale neural network that combines several hierarchical concepts within a single estimation framework. These concepts include (i) a partially shared coarse-to-fine architecture, (ii) multi-scale features, (iii) a hierarchical cost volume and (iv) a multi-scale multi-iteration loss. Experiments on MPI Sintel and KITTI clearly demonstrate the benefits of our approach. They show not only substantial improvements compared to RAFT, but also state-of-the-art results -- in particular in non-occluded regions.
@inproceedings{Jahedi2022_ICIP,
abstract = {Many classical and learning-based optical flow methods rely on hierarchical concepts to improve both accuracy and robustness. However, one of the currently most successful approaches -- RAFT -- hardly exploits such concepts. In this work, we show that multi-scale ideas are still valuable. More precisely, using RAFT as a baseline, we propose a novel multi-scale neural network that combines several hierarchical concepts within a single estimation framework. These concepts include (i) a partially shared coarse-to-fine architecture, (ii) multi-scale features, (iii) a hierarchical cost volume and (iv) a multi-scale multi-iteration loss. Experiments on MPI Sintel and KITTI clearly demonstrate the benefits of our approach. They show not only substantial improvements compared to RAFT, but also state-of-the-art results -- in particular in non-occluded regions.},
added-at = {2022-08-04T10:22:17.000+0200},
author = {Jahedi, Azin and Mehl, Lukas and Rivinius, Marc and Bruhn, Andrés},
biburl = {https://puma.ub.uni-stuttgart.de/bibtex/23ca6ab269f14747716126e8ce143a2bd/christinawarren},
booktitle = {IEEE International Conference on Image Processing (ICIP)},
doi = {10.48550/arXiv.2207.12163},
interhash = {e50b9467262eb8c0b64043ce9647ed53},
intrahash = {3ca6ab269f14747716126e8ce143a2bd},
keywords = {2022 B04 sfbtrr161 visus visus:bruhnas visus:mehlls},
timestamp = {2022-08-04T08:22:17.000+0200},
title = {Multi-Scale {RAFT}: combining hierarchical concepts for learning-based optical flow estimation},
url = {/brokenurl# https://doi.org/10.48550/arXiv.2207.12163},
year = 2022
}