`skimage.data`#

Example images and datasets.

A curated set of general purpose and scientific images used in tests, examples, and documentation.

Newer datasets are no longer included as part of the package, but are downloaded on demand. To make data available offline, use download_all().

`astronaut`	Color image of the astronaut Eileen Collins.
`binary_blobs`	Generate synthetic binary image with several rounded blob-like objects.
`brain`	Subset of data from the University of North Carolina Volume Rendering Test Data Set.
`brick`	Brick wall.
`camera`	Gray-level "camera" image.
`cat`	Chelsea the cat.
`cell`	Cell floating in saline.
`cells3d`	3D fluorescence microscopy image of cells.
`checkerboard`	Checkerboard image.
`chelsea`	Chelsea the cat.
`clock`	Motion blurred clock.
`coffee`	Coffee cup.
`coins`	Greek coins from Pompeii.
`colorwheel`	Color Wheel.
`download_all`	Download all datasets for use with scikit-image offline.
`eagle`	A golden eagle.
`file_hash`	Calculate the hash of a given file.
`grass`	Grass.
`gravel`	Gravel
`horse`	Black and white silhouette of a horse.
`hubble_deep_field`	Hubble eXtreme Deep Field.
`human_mitosis`	Image of human cells undergoing mitosis.
`immunohistochemistry`	Immunohistochemical (IHC) staining with hematoxylin counterstaining.
`kidney`	Mouse kidney tissue.
`lbp_frontal_face_cascade_filename`	Return the path to the XML file containing the weak classifier cascade.
`lfw_subset`	Subset of data from the LFW dataset.
`lily`	Lily of the valley plant stem.
`logo`	Scikit-image logo, a RGBA image.
`microaneurysms`	Gray-level "microaneurysms" image.
`moon`	Surface of the moon.
`nickel_solidification`	Image sequence of synchrotron x-radiographs showing the rapid solidification of a nickel alloy sample.
`page`	Scanned page.
`palisades_of_vogt`	Return image sequence of in-vivo tissue showing the palisades of Vogt.
`protein_transport`	Microscopy image sequence with fluorescence tagging of proteins re-localizing from the cytoplasmic area to the nuclear envelope.
`retina`	Human retina.
`rocket`	Launch photo of DSCOVR on Falcon 9 by SpaceX.
`shepp_logan_phantom`	Shepp Logan Phantom.
`skin`	Microscopy image of dermis and epidermis (skin layers).
`stereo_motorcycle`	Rectified stereo image pair with ground-truth disparities.
`text`	Gray-level "text" image used for corner detection.
`vortex`	Case B1 image pair from the first PIV challenge.

skimage.data.astronaut()[source]#

Color image of the astronaut Eileen Collins.

Photograph of Eileen Collins, an American astronaut. She was selected as an astronaut in 1992 and first piloted the space shuttle STS-63 in 1995. She retired in 2006 after spending a total of 38 days, 8 hours and 10 minutes in outer space.

This image was downloaded from the NASA Great Images database <https://flic.kr/p/r9qvLn>`__.

No known copyright restrictions, released into the public domain.

Returns:

astronaut(512, 512, 3) uint8 ndarray: Astronaut image.

General-purpose images

Block views on images/arrays

RGB to grayscale

Adapting gray-scale filters to RGB images

Active Contour Model

Rescale, resize, and downscale

Build image pyramids

Piecewise Affine Transformation

Image Deconvolution

Using window functions with images

Image Deconvolution

Estimate strength of blur

Fill in defects with inpainting

Non-local means denoising for preserving textures

Histogram of Oriented Gradients

CENSURE feature detector

ORB feature detector and binary descriptor

Gabors / Primary Visual Cortex “Simple Cells” from an Image

Gabors / Primary Visual Cortex "Simple Cells" from an Image

BRIEF binary descriptor

SIFT feature detector and descriptor extractor

Comparison of segmentation and superpixel algorithms

Flood Fill

Face detection using a cascade classifier

skimage.data.binary_blobs(length=512, blob_size_fraction=0.1, n_dim=2, volume_fraction=0.5, rng=None)[source]#

Generate synthetic binary image with several rounded blob-like objects.

Parameters:

lengthint, optional: Linear size of output image.
blob_size_fractionfloat, optional: Typical linear size of blob, as a fraction of length, should be smaller than 1.
n_dimint, optional: Number of dimensions of output image.
volume_fractionfloat, default 0.5: Fraction of image pixels covered by the blobs (where the output is 1). Should be in [0, 1].
rng{numpy.random.Generator, int}, optional: Pseudo-random number generator. By default, a PCG64 generator is used (see numpy.random.default_rng()). If rng is an int, it is used to seed the generator.

Returns:

blobsndarray of bools: Output binary image

Examples

>>> from skimage import data
>>> data.binary_blobs(length=5, blob_size_fraction=0.2)  
array([[ True, False,  True,  True,  True],
       [ True,  True,  True, False,  True],
       [False,  True, False,  True,  True],
       [ True, False, False,  True,  True],
       [ True, False, False, False,  True]])
>>> blobs = data.binary_blobs(length=256, blob_size_fraction=0.1)
>>> # Finer structures
>>> blobs = data.binary_blobs(length=256, blob_size_fraction=0.05)
>>> # Blobs cover a smaller volume fraction of the image
>>> blobs = data.binary_blobs(length=256, volume_fraction=0.3)

General-purpose images

Skeletonize

Random walker segmentation

Explore and visualize region properties with pandas

Colocalization metrics

skimage.data.brain()[source]#

Subset of data from the University of North Carolina Volume Rendering Test Data Set.

The full dataset is available at [1].

Returns:

image(10, 256, 256) uint16 ndarray

Notes

The 3D volume consists of 10 layers from the larger volume.

References

[1]

https://graphics.stanford.edu/data/voldata/

Local Histogram Equalization

Rank filters

skimage.data.brick()[source]#

Brick wall.

Returns:

brick(512, 512) uint8 image: A small section of a brick wall.

Notes

The original image was downloaded from CC0Textures and licensed under the Creative Commons CC0 License.

A perspective transform was then applied to the image, prior to rotating it by 90 degrees, cropping and scaling it to obtain the final image.

General-purpose images

Gabor filter banks for texture classification

Local Binary Pattern for texture classification

skimage.data.camera()[source]#

Gray-level “camera” image.

Can be used for segmentation and denoising examples.

Returns:

camera(512, 512) uint8 ndarray: Camera image.

Notes

No copyright restrictions. CC0 by the photographer (Lav Varshney).

Changed in version 0.18: This image was replaced due to copyright restrictions. For more information, please see [1].

References

[1]

scikit-image/scikit-image#3927

General-purpose images

Using simple NumPy operations for manipulating images

Tinting gray-scale images

Straight line Hough transform

Edge operators

Structural similarity index

Image Registration

Masked Normalized Cross-Correlation

Entropy

Band-pass filtering by Difference of Gaussians

Butterworth Filters

Dense DAISY feature description

GLCM Texture Features

Thresholding

Chan-Vese Segmentation

Multi-Otsu Thresholding

skimage.data.cat()[source]#

Chelsea the cat.

An example with texture, prominent edges in horizontal and diagonal directions, as well as features of differing scales.

Returns:

chelsea(300, 451, 3) uint8 ndarray: Chelsea image.

Notes

No copyright restrictions. CC0 by the photographer (Stefan van der Walt).

General-purpose images

Render text onto an image

skimage.data.cell()[source]#

Cell floating in saline.

This is a quantitative phase image retrieved from a digital hologram using the Python library qpformat. The image shows a cell with high phase value, above the background phase.

Because of a banding pattern artifact in the background, this image is a good test of thresholding algorithms. The pixel spacing is 0.107 µm.

These data were part of a comparison between several refractive index retrieval techniques for spherical objects as part of [1].

This image is CC0, dedicated to the public domain. You may copy, modify, or distribute it without asking permission.

Returns:

cell(660, 550) uint8 array: Image of a cell.

References

[1]

Paul Müller, Mirjam Schürmann, Salvatore Girardo, Gheorghe Cojoc, and Jochen Guck. “Accurate evaluation of size and refractive index for spherical objects in quantitative phase imaging.” Optics Express 26(8): 10729-10743 (2018). DOI:10.1364/OE.26.010729

Li thresholding

skimage.data.cells3d()[source]#

3D fluorescence microscopy image of cells.

The returned data is a 3D multichannel array with dimensions provided in (z, c, y, x) order. Each voxel has a size of (0.29 0.26 0.26) micrometer. Channel 0 contains cell membranes, channel 1 contains nuclei.

Returns:

cells3d: (60, 2, 256, 256) uint16 ndarray: The volumetric images of cells taken with an optical microscope.

Notes

The data for this was provided by the Allen Institute for Cell Science.

It has been downsampled by a factor of 4 in the row and column dimensions to reduce computational time.

The microscope reports the following voxel spacing in microns:

Original voxel size is (0.290, 0.065, 0.065).

Scaling factor is (1, 4, 4) in each dimension.

After rescaling the voxel size is (0.29 0.26 0.26).

Datasets with 3 or more spatial dimensions

3D adaptive histogram equalization

Use rolling-ball algorithm for estimating background intensity

Explore 3D images (of cells)

skimage.data.checkerboard()[source]#

Checkerboard image.

Checkerboards are often used in image calibration, since the corner-points are easy to locate. Because of the many parallel edges, they also visualise distortions particularly well.

Returns:

checkerboard(200, 200) uint8 ndarray: Checkerboard image.

General-purpose images

Swirl

Use thin-plate splines for image warping

Robust matching using RANSAC

Corner detection

Flood Fill

skimage.data.chelsea()[source]#

Chelsea the cat.

An example with texture, prominent edges in horizontal and diagonal directions, as well as features of differing scales.

Returns:

chelsea(300, 451, 3) uint8 ndarray: Chelsea image.

Notes

No copyright restrictions. CC0 by the photographer (Stefan van der Walt).

Histogram matching

Types of homographies

Calibrating Denoisers Using J-Invariance

Denoising a picture

Shift-invariant wavelet denoising

Phase Unwrapping

Wavelet denoising

Full tutorial on calibrating Denoisers Using J-Invariance

Flood Fill

skimage.data.clock()[source]#

Motion blurred clock.

This photograph of a wall clock was taken while moving the camera in an approximately horizontal direction. It may be used to illustrate inverse filters and deconvolution.

Released into the public domain by the photographer (Stefan van der Walt).

Returns:

clock(300, 400) uint8 ndarray: Clock image.

General-purpose images

skimage.data.coffee()[source]#

Coffee cup.

This photograph is courtesy of Pikolo Espresso Bar. It contains several elliptical shapes as well as varying texture (smooth porcelain to coarse wood grain).

Returns:

coffee(400, 600, 3) uint8 ndarray: Coffee image.

Notes

No copyright restrictions. CC0 by the photographer (Rachel Michetti).

RGB to HSV

Histogram matching

Circular and Elliptical Hough Transforms

Region Boundary based Region adjacency graphs (RAGs)

Region adjacency graph (RAG) Thresholding

Normalized Cut

Drawing Region Adjacency Graphs (RAGs)

Region adjacency graph (RAG) Merging

Hierarchical Merging of Region Boundary RAGs

skimage.data.coins()[source]#

Greek coins from Pompeii.

This image shows several coins outlined against a gray background. It is especially useful in, e.g. segmentation tests, where individual objects need to be identified against a background. The background shares enough grey levels with the coins that a simple segmentation is not sufficient.

Returns:

coins(303, 384) uint8 ndarray: Coins image.

Notes

This image was downloaded from the Brooklyn Museum Collection.

No known copyright restrictions.

Filtering regional maxima

Circular and Elliptical Hough Transforms

Hysteresis thresholding

Mean filters

Template Matching

Multi-Block Local Binary Pattern for texture classification

Sliding window histogram

Find Regular Segments Using Compact Watershed

Finding local maxima

Expand segmentation labels without overlap

Label image regions

Find the intersection of two segmentations

Morphological Snakes

Measure region properties

Evaluating segmentation metrics

Use rolling-ball algorithm for estimating background intensity

Visual image comparison

Comparing edge-based and region-based segmentation

skimage.data.colorwheel()[source]#

Color Wheel.

Returns:

colorwheel(370, 371, 3) uint8 image: A colorwheel.

General-purpose images

skimage.data.download_all(directory=None)[source]#

Download all datasets for use with scikit-image offline.

Scikit-image datasets are no longer shipped with the library by default. This allows us to use higher quality datasets, while keeping the library download size small.

This function requires the installation of an optional dependency, pooch, to download the full dataset. Follow installation instruction found at

https://scikit-image.org/docs/stable/user_guide/install.html

Call this function to download all sample images making them available offline on your machine.

Parameters:

directory: path-like, optional: The directory where the dataset should be stored.

Raises:

ModuleNotFoundError:: If pooch is not install, this error will be raised.

Notes

scikit-image will only search for images stored in the default directory. Only specify the directory if you wish to download the images to your own folder for a particular reason. You can access the location of the default data directory by inspecting the variable skimage.data.data_dir.

skimage.data.eagle()[source]#

A golden eagle.

Suitable for examples on segmentation, Hough transforms, and corner detection.

Returns:

eagle(2019, 1826) uint8 ndarray: Eagle image.

Notes

No copyright restrictions. CC0 by the photographer (Dayane Machado).

Markers for watershed transform

skimage.data.file_hash(fname, alg='sha256')[source]#

Calculate the hash of a given file.

Useful for checking if a file has changed or been corrupted.

Parameters:

fnamestr: The name of the file.
algstr: The type of the hashing algorithm

Returns:

hashstr: The hash of the file.

Examples

>>> fname = "test-file-for-hash.txt"
>>> with open(fname, "w") as f:
...     __ = f.write("content of the file")
>>> print(file_hash(fname))
0fc74468e6a9a829f103d069aeb2bb4f8646bad58bf146bb0e3379b759ec4a00
>>> import os
>>> os.remove(fname)

skimage.data.grass()[source]#

Grass.

Returns:

grass(512, 512) uint8 image: Some grass.

Notes

The original image was downloaded from DeviantArt and licensed under the Creative Commons CC0 License.

The downloaded image was cropped to include a region of (512, 512) pixels around the top left corner, converted to grayscale, then to uint8 prior to saving the result in PNG format.

Gabor filter banks for texture classification

Local Binary Pattern for texture classification

skimage.data.gravel()[source]#

Gravel

Returns:

gravel(512, 512) uint8 image: Grayscale gravel sample.

Notes

The original image was downloaded from CC0Textures and licensed under the Creative Commons CC0 License.

The downloaded image was then rescaled to (1024, 1024), then the top left (512, 512) pixel region was cropped prior to converting the image to grayscale and uint8 data type. The result was saved using the PNG format.

Band-pass filtering by Difference of Gaussians

Gabor filter banks for texture classification

Local Binary Pattern for texture classification

skimage.data.horse()[source]#

Black and white silhouette of a horse.

This image was downloaded from openclipart

No copyright restrictions. CC0 given by owner (Andreas Preuss (marauder)).

Returns:

horse(328, 400) bool ndarray: Horse image.

Convex Hull

Skeletonize

Morphological Filtering

skimage.data.hubble_deep_field()[source]#

Hubble eXtreme Deep Field.

This photograph contains the Hubble Telescope’s farthest ever view of the universe. It can be useful as an example for multi-scale detection.

Returns:

hubble_deep_field(872, 1000, 3) uint8 ndarray: Hubble deep field image.

Notes

This image was downloaded from HubbleSite.

The image was captured by NASA and may be freely used in the public domain.

Scientific images

Removing small objects in grayscale images with a top hat filter

Full tutorial on calibrating Denoisers Using J-Invariance

Removing objects

Blob Detection

Extrema

skimage.data.human_mitosis()[source]#

Image of human cells undergoing mitosis.

Returns:

human_mitosis: (512, 512) uint8 ndarray: Data of human cells undergoing mitosis taken during the preparation of the manuscript in [1].

Notes

Copyright David Root. Licensed under CC-0 [2].

References

[1]

Moffat J, Grueneberg DA, Yang X, Kim SY, Kloepfer AM, Hinkle G, Piqani B, Eisenhaure TM, Luo B, Grenier JK, Carpenter AE, Foo SY, Stewart SA, Stockwell BR, Hacohen N, Hahn WC, Lander ES, Sabatini DM, Root DE (2006) A lentiviral RNAi library for human and mouse genes applied to an arrayed viral high-content screen. Cell, 124(6):1283-98 / :DOI: 10.1016/j.cell.2006.01.040 PMID 16564017

[2]

GitHub licensing discussion CellProfiler/examples#41

Segment human cells (in mitosis)

skimage.data.immunohistochemistry()[source]#

Immunohistochemical (IHC) staining with hematoxylin counterstaining.

This picture shows colonic glands where the IHC expression of FHL2 protein is revealed with DAB. Hematoxylin counterstaining is applied to enhance the negative parts of the tissue.

This image was acquired at the Center for Microscopy And Molecular Imaging (CMMI).

No known copyright restrictions.

Returns:

immunohistochemistry(512, 512, 3) uint8 ndarray: Immunohistochemistry image.

Scientific images

Separate colors in immunohistochemical staining

Apply maskSLIC vs SLIC

skimage.data.kidney()[source]#

Mouse kidney tissue.

This biological tissue on a pre-prepared slide was imaged with confocal fluorescence microscopy (Nikon C1 inverted microscope). Image shape is (16, 512, 512, 3). That is 512x512 pixels in X-Y, 16 image slices in Z, and 3 color channels (emission wavelengths 450nm, 515nm, and 605nm, respectively). Real-space voxel size is 1.24 microns in X-Y, and 1.25 microns in Z. Data type is unsigned 16-bit integers.

Returns:

kidney(16, 512, 512, 3) uint16 ndarray: Kidney 3D multichannel image.

Notes

This image was acquired by Genevieve Buckley at Monasoh Micro Imaging in 2018. License: CC0

Interact with 3D images (of kidney tissue)

Estimate anisotropy in a 3D microscopy image

skimage.data.lbp_frontal_face_cascade_filename()[source]#

Return the path to the XML file containing the weak classifier cascade.

These classifiers were trained using LBP features. The file is part of the OpenCV repository [1].

References

[1]

OpenCV lbpcascade trained files opencv/opencv

Face detection using a cascade classifier

skimage.data.lfw_subset()[source]#

Subset of data from the LFW dataset.

This database is a subset of the LFW database containing:

100 faces
100 non-faces

The full dataset is available at [2].

Returns:

images(200, 25, 25) uint8 ndarray: 100 first images are faces and subsequent 100 are non-faces.

Notes

The faces were randomly selected from the LFW dataset and the non-faces were extracted from the background of the same dataset. The cropped ROIs have been resized to a 25 x 25 pixels.

References

[1]

Huang, G., Mattar, M., Lee, H., & Learned-Miller, E. G. (2012). Learning to align from scratch. In Advances in Neural Information Processing Systems (pp. 764-772).

[2]

http://vis-www.cs.umass.edu/lfw/

Specific images

Face classification using Haar-like feature descriptor

skimage.data.lily()[source]#

Lily of the valley plant stem.

This plant stem on a pre-prepared slide was imaged with confocal fluorescence microscopy (Nikon C1 inverted microscope). Image shape is (922, 922, 4). That is 922x922 pixels in X-Y, with 4 color channels. Real-space voxel size is 1.24 microns in X-Y. Data type is unsigned 16-bit integers.

Returns:

lily(922, 922, 4) uint16 ndarray: Lily 2D multichannel image.

Notes

This image was acquired by Genevieve Buckley at Monasoh Micro Imaging in 2018. License: CC0

Scientific images

skimage.data.logo()[source]#

Scikit-image logo, a RGBA image.

Returns:

logo(500, 500, 4) uint8 ndarray: Logo image.

skimage.data.microaneurysms()[source]#

Gray-level “microaneurysms” image.

Detail from an image of the retina (green channel). The image is a crop of image 07_dr.JPG from the High-Resolution Fundus (HRF) Image Database: https://www5.cs.fau.de/research/data/fundus-images/

Returns:

microaneurysms(102, 102) uint8 ndarray: Retina image with lesions.

Notes

No copyright restrictions. CC0 given by owner (Andreas Maier).

References

[1]

Budai, A., Bock, R, Maier, A., Hornegger, J., Michelson, G. (2013). Robust Vessel Segmentation in Fundus Images. International Journal of Biomedical Imaging, vol. 2013, 2013. DOI:10.1155/2013/154860

Scientific images

Attribute operators

skimage.data.moon()[source]#

Surface of the moon.

This low-contrast image of the surface of the moon is useful for illustrating histogram equalization and contrast stretching.

Returns:

moon(512, 512) uint8 ndarray: Moon image.

Scientific images

Gamma and log contrast adjustment

Histogram Equalization

Local Histogram Equalization

Assemble images with simple image stitching

Unsharp masking

Filling holes and finding peaks

skimage.data.nickel_solidification()[source]#

Image sequence of synchrotron x-radiographs showing the rapid solidification of a nickel alloy sample.

Returns:

nickel_solidification: (11, 384, 512) uint16 ndarray

Notes

See info under nickel_solidification.tif at scikit-image/data/-/blob/master/README.md#data.

Track solidification of a metallic alloy

skimage.data.page()[source]#

Scanned page.

This image of printed text is useful for demonstrations requiring uneven background illumination.

Returns:

page(191, 384) uint8 ndarray: Page image.

Attribute operators

Thresholding

Niblack and Sauvola Thresholding

Use rolling-ball algorithm for estimating background intensity

Thresholding

Rank filters

skimage.data.palisades_of_vogt()[source]#

Return image sequence of in-vivo tissue showing the palisades of Vogt.

In the human eye, the palisades of Vogt are normal features of the corneal limbus, which is the border between the cornea and the sclera (i.e., the white of the eye). In the image sequence, there are some dark spots due to the presence of dust on the reference mirror.

Returns:

palisades_of_vogt: (60, 1440, 1440) uint16 ndarray

Notes

See info under in-vivo-cornea-spots.tif at scikit-image/data/-/blob/master/README.md#data.

Restore spotted cornea image with inpainting

skimage.data.protein_transport()[source]#

Microscopy image sequence with fluorescence tagging of proteins re-localizing from the cytoplasmic area to the nuclear envelope.

Returns:

protein_transport: (15, 2, 180, 183) uint8 ndarray

Notes

See info under NPCsingleNucleus.tif at scikit-image/data/-/blob/master/README.md#data.

Colocalization metrics

Measure fluorescence intensity at the nuclear envelope

skimage.data.retina()[source]#

Human retina.

This image of a retina is useful for demonstrations requiring circular images.

Returns:

retina(1411, 1411, 3) uint8 ndarray: Retina image in RGB.

Notes

This image was downloaded from wikimedia. This file is made available under the Creative Commons CC0 1.0 Universal Public Domain Dedication.

References

[1]

Häggström, Mikael (2014). “Medical gallery of Mikael Häggström 2014”. WikiJournal of Medicine 1 (2). DOI:10.15347/wjm/2014.008. ISSN 2002-4436. Public Domain

Scientific images

Ridge operators

Using Polar and Log-Polar Transformations for Registration

Use pixel graphs to find an object’s geodesic center

Use pixel graphs to find an object's geodesic center

skimage.data.rocket()[source]#

Launch photo of DSCOVR on Falcon 9 by SpaceX.

This is the launch photo of Falcon 9 carrying DSCOVR lifted off from SpaceX’s Launch Complex 40 at Cape Canaveral Air Force Station, FL.

Returns:

rocket(427, 640, 3) uint8 ndarray: Rocket image.

Notes

This image was downloaded from SpaceX Photos.

The image was captured by SpaceX and released in the public domain.

skimage.data.shepp_logan_phantom()[source]#

Shepp Logan Phantom.

Returns:

phantom(400, 400) float64 image: Image of the Shepp-Logan phantom in grayscale.

References

[1]

L. A. Shepp and B. F. Logan, “The Fourier reconstruction of a head section,” in IEEE Transactions on Nuclear Science, vol. 21, no. 3, pp. 21-43, June 1974. DOI:10.1109/TNS.1974.6499235

Scientific images

Radon transform

Morphological Filtering

skimage.data.skin()[source]#

Microscopy image of dermis and epidermis (skin layers).

Hematoxylin and eosin stained slide at 10x of normal epidermis and dermis with a benign intradermal nevus.

Returns:

skin(960, 1280, 3) RGB image of uint8

Notes

This image requires an Internet connection the first time it is called, and to have the pooch package installed, in order to fetch the image file from the scikit-image datasets repository.

The source of this image is https://en.wikipedia.org/wiki/File:Normal_Epidermis_and_Dermis_with_Intradermal_Nevus_10x.JPG

The image was released in the public domain by its author Kilbad.

Scientific images

Trainable segmentation using local features and random forests

skimage.data.stereo_motorcycle()[source]#

Rectified stereo image pair with ground-truth disparities.

The two images are rectified such that every pixel in the left image has its corresponding pixel on the same scanline in the right image. That means that both images are warped such that they have the same orientation but a horizontal spatial offset (baseline). The ground-truth pixel offset in column direction is specified by the included disparity map.

The two images are part of the Middlebury 2014 stereo benchmark. The dataset was created by Nera Nesic, Porter Westling, Xi Wang, York Kitajima, Greg Krathwohl, and Daniel Scharstein at Middlebury College. A detailed description of the acquisition process can be found in [1].

The images included here are down-sampled versions of the default exposure images in the benchmark. The images are down-sampled by a factor of 4 using the function skimage.transform.downscale_local_mean. The calibration data in the following and the included ground-truth disparity map are valid for the down-sampled images:

Focal length:           994.978px
Principal point x:      311.193px
Principal point y:      254.877px
Principal point dx:      31.086px
Baseline:               193.001mm

Returns:

img_left(500, 741, 3) uint8 ndarray: Left stereo image.
img_right(500, 741, 3) uint8 ndarray: Right stereo image.
disp(500, 741, 3) float ndarray: Ground-truth disparity map, where each value describes the offset in column direction between corresponding pixels in the left and the right stereo images. E.g. the corresponding pixel of img_left[10, 10 + disp[10, 10]] is img_right[10, 10]. NaNs denote pixels in the left image that do not have ground-truth.

Notes

The original resolution images, images with different exposure and lighting, and ground-truth depth maps can be found at the Middlebury website [2].

References

[1]

D. Scharstein, H. Hirschmueller, Y. Kitajima, G. Krathwohl, N. Nesic, X. Wang, and P. Westling. High-resolution stereo datasets with subpixel-accurate ground truth. In German Conference on Pattern Recognition (GCPR 2014), Muenster, Germany, September 2014.

[2]

http://vision.middlebury.edu/stereo/data/scenes2014/

Specific images