Illaitar/stereo-color-sharpness-mismatch

Real-World Stereo Color and Sharpness Mismatch Dataset

Project page · Paper (arXiv:2303.06657) · Code

A real-world dataset of stereoscopic video with real color and sharpness mismatches between views, built for evaluating color-mismatch-correction methods. Unlike earlier datasets for this task, the distortions are not synthesized: they come from a beam-splitter rig.

The scenes were filmed with a beam splitter and three cameras. The beam splitter sets a zero stereobase between the left camera and the left ground-truth camera. This gives a pixel-aligned pair of a distorted left view and a ground-truth left view. The right view has the same color and sharpness characteristics as the ground-truth left view. Scenes include moving objects, light sources, complex textures and glares.

Postprocessing: the horizontally flipped left distorted view was matched to the left ground-truth view, and the stereopair was rectified with a homography estimated by MAGSAC++ (SIFT + brute-force matching for the left views, LoFTR for rectification). A Monge–Kantorovich linear colour mapping corrected the small loss in the transmitted beam. Temporal alignment was done using the cameras' audio streams. The postprocessing scripts are in the code repository.

Task

Given left_distorted and the reference right view, restore left_gt: transfer the colors (and sharpness) of the right view onto the corresponding regions of the distorted left view.

Dataset structure

Each row is one stereo frame:

Column	Type	Description
left_distorted	image	Left view with real color/sharpness mismatch (file suffix _LD)
left_gt	image	Ground-truth left view, pixel-aligned with left_distorted (suffix _L)
right	image	Right view, the reference for correction (suffix _R)
scene	string	Scene name; a -sharpness suffix marks scenes that also have a sharpness mismatch
sharpness_mismatch	bool	True if the scene name ends with -sharpness
scene_frame	int32	Frame index within the scene (0–49)
frame	int32	Frame index within the split; matches the file number in the original archive

Splits

Split	Scenes	Frames	Resolution
train	18	900	960×720
validation	3	150	960×720
test	3	150	512×512

Every scene has 50 consecutive frames. Images are the original lossless PNGs from the release archive, stored as-is.

Split	Scenes
train	green-disk, green-disk-sharpness, green-warrior, green-warrior-sharpness, large-glass, large-glass-sharpness, metal-ball-sharpness, moving-ball-sharpness, moving-cradle-sharpness, photo-card, photo-card-sharpness, plasma-boy-sharpness, red-dice, red-dice-sharpness, red-pattern, red-pattern-sharpness, shimmering-ball, shimmering-ball-sharpness
validation	color-pattern-sharpness, color-gradient-sharpness, lego-tower-sharpness
test	fluffy-toy, color-gradient, color-pattern

The original archive does not contain per-frame scene labels; info.csv lists the scenes of each split in order. scene and scene_frame are derived from it as frame // 50 and frame % 50. The boundaries were checked against the frame content.

Original archive

The release archive is in original/stereo-color-sharpness-mismatch.zip, unchanged: folders Train/, Validation/, Test/ with files NNNN_L.png, NNNN_LD.png, NNNN_R.png, plus info.csv.

Usage

from datasets import load_dataset

ds = load_dataset("Illaitar/stereo-color-sharpness-mismatch")
sample = ds["test"][0]
sample["left_distorted"], sample["left_gt"], sample["right"]  # PIL images

Authors

Egor Chistov, Nikita Alutis, Maxim Velikanov, Dmitriy Vatolin. Video Processing Group, MSU Graphics & Media Lab.

Questions: stereo-mismatch-dataset@videoprocessing.ai

Citation

@article{chistov2023color,
  title={Color Mismatches in Stereoscopic Video: Real-World Dataset and Deep Correction Method},
  author={Chistov, Egor and Alutis, Nikita and Velikanov, Maxim and Vatolin, Dmitriy},
  howpublished={arXiv preprint arXiv:2303.06657},
  year={2023}
}