🚨 This repository contains download links to our dataset, code snippets, and trained deep stereo models of our work "FlowSeek: Optical Flow Made Easier with Depth Foundation Models and Motion Bases", ICCV 2025

by Matteo Poggi¹ and Fabio Tosi¹

University of Bologna¹

We present FlowSeek, a novel framework for optical flow requiring minimal hardware resources for training. FlowSeek marries the latest advances on the design space of optical flow networks with cutting-edge single-image depth foundation models and classical low-dimensional motion parametrization, implementing a compact, yet accurate architecture. FlowSeek is trained on a single consumer-grade GPU, a hardware budget about 8× lower compared to most recent methods, and still achieves superior cross-dataset generalization on Sintel Final and KITTI, with a relative improvement of 10 and 15% over the previous state-of-the-art SEA-RAFT, as well as on Spring and LayeredFlow datasets.

Contributions:

• We introduce FlowSeek, the first optical flow model that integrates a pre-trained depth foundation model.

• We explore different design strategies to best exploit the prior knowledge of the foundation model for the optical flow estimation task.

• We develop several variants of FlowSeek, implementing different trade-offs between accuracy and efficiency, yet maintaining the single-GPU requirement at training time.

🖋️ If you find this code useful in your research, please cite:

@InProceedings{Poggi_2025_ICCV,
    author    = {Poggi, Matteo and Tosi, Fabio},
    title     = {FlowSeek: Optical Flow Made Easier with Depth Foundation Models and Motion Bases},
    booktitle = {Proceedings of the International Conference on Computer Vision (ICCV)},
    year      = {2025},
}

This code has been tested with nVIDIA Drivers version 550.54.14 and CUDA Libraries version 12.1.

1. Dependencies: Ensure that you have installed all the necessary dependencies. The list of dependencies can be found in the requirements.txt file.

2. Download required datasets: To evaluate/train FlowSeek, you will need to download the required datasets: FlyingChairs, FlyingThings3D, Sintel, KITTI, HD1K, TartanAir, and Spring. By default datasets.py will search for the datasets in the locations pointed by the config file configs/datapaths.py. You can edit the file with the proper paths in your file system

3. Download pre-trained models: You can download Depth Anything v2 pre-trained weights (and ours) by running

bash scripts/get_weights.sh

You can train FlowSeek by setting the variable SIZE either to T or M inside scripts/train.sh, and then run

bash scripts/train.sh

You can evaluate our pre-trained models by simply running

bash scripts/eval.sh

or you can evaluate your own checkpoint $checkpoint by running

python evaluate.py --cfg config/eval/flowseek-[T,S,M,L].json --model $checkpoint --dataset [sintel, kitti, spring, layeredflow]

You can reproduce the qualitative results from either the main paper or the supplementary material by running

bash scripts/qualitatives.sh

For questions, please send an email to m.poggi@unibo.it or fabio.tosi5@unibo.it

We would like to extend our sincere appreciation to the authors of the following projects for making their code available, which we have utilized in our work:

• We would like to thank the authors of SEA-RAFT and Depth Anything v2 for providing their code, which has been instrumental to implement FlowSeek

• We would also like to thanks the authors of DPT, Depth Anything, FlowFormer and CRAFT for providing their code, which has been instrumental in our ablation studies.