RobustSpring: Benchmarking Robustness to Image Corruptions for Optical Flow, Scene Flow and Stereo

This dataset provides structured metadata only for the RobustSpring dataset. All image samples are referenced by relative file paths, and must be paired with local image data downloaded separately from the public release site.

• Dataset on the Hub: jeschmalfuss/RobustSpring
• Image Data: RobustSpring

For the related research see

RobustSpring: Benchmarking Robustness to Image Corruptions for Optical Flow, Scene Flow and Stereo
Jenny Schmalfuss*, Victor Oei*, Lukas Mehl, Madlen Bartsch, Shashank Agnihotri, Margret Keuper, Andrés Bruhn
https://doi.org/10.48550/arXiv.2505.09368

RobustSpring is an image-corruption dataset for optical flow, scene flow and stereo, that applies 20 different image corruption to the test split of the Spring dataset. The combined Spring and RobustSpring website is at spring-benchmark.org

---

Dataset Overview

Each sample in this dataset represents one data sample on which to predict:

• Optical Flow
• Scene Flow
• Stereo Disparity

The dataset contains only file paths to local image files. The raw image data must be downloaded separately.

---

Download Image Data

Please download the raw image data zips files from:

https://doi.org/10.18419/DARUS-5047

After downloading:

1. Extract all contents to a local data/ folder.
2. Ensure the folder structure looks like:

/data/
  brightness/
    test/
      scene_0003/
        frame_left/
          frame_left_0001.png
          frame_left_0002.png
          ...
        frame_right/
          frame_right_0001.png
          frame_right_0002.png
          ...
      scene_0019/
        frame_left/
          ...
        frame_right/
          ...
      scene_0028
      ...
  contrast/
    test/
      scene_0003/
      scene_0019/
      ...
  defocus_blur/
    test/
      scene_0003/
      scene_0019/
      ...
  ...

---

Dataset Structure

Each sample in the dataset includes:

Field	Type	Description
sample_type	string	"optic-flow", "scene-flow" or "stereo"
corruption	string	Image corruption type
split	string	Dataset split. test for all data.
scene_id	string	Spring's scene ID
frame_leftright	string	If data is centered on left or right stereo frame
frame_forwardbackward	string	For optic- and scene-flow. Forward or backward in time.
index	int32	Data sample index. Own indices for optical flow, scene flow and stereo.
image1_path	string	Relative path to pivot image
image2_path	string	Relative path to pivot image at next time step (OF & SF only)
image1s_path	string	Relative path to stereo of pivot image (SF and S only)
image2s_path	string	Relative path to stereo of image at next time step (SF only)

No image content is stored. Paths only.

---

How to Use

1. Install Dependencies

pip install datasets Pillow

2. Load the Dataset

from datasets import load_dataset

dataset = load_dataset("jeschmalfuss/RobustSpring", split="test")  # all samples

3. Filtering by Data Type

You can filter the dataset to only retrieve the type of samples you're interested in: optical flow, scene flow or stereo.

dataset_optic_flow = dataset.filter(lambda x: x["sample_type"] == "optic-flow")
dataset_scene_flow = dataset.filter(lambda x: x["sample_type"] == "scene-flow")
dataset_stereo     = dataset.filter(lambda x: x["sample_type"] == "stereo")

4. Set Local Path to Images

import os
from PIL import Image

base_path = "/absolute/path/to/data"  # where you extracted the downloaded zip

sample = dataset_optic_flow[0]
img1 = Image.open(os.path.join(base_path, sample["image1_path"]))
img2 = Image.open(os.path.join(base_path, sample["image2_path"]))

---

License

The RobustSpring dataset is licensed under CC-BY-4.0