日本語 · English
User note (2026-08-16): “There must be plenty of source papers and GitHub repos behind these.” Following Fullseye’s public-disclosure policy ([[project_imgevolve_goal_knowledge_layer_2026_08_13]]), we record the provenance, license, source paper, and public repository of the collected sample images in an auditable form. We do not download any images from external sources without permission (synthetic = our own work; everything else = only the classic images bundled with scikit-image). The MVTec/HALCON sample images are proprietary, so we do not collect them.
Collection = studio_assets/sample_images/ (machine-readable provenance is in manifest.json in the same dir).
Regenerate = py -3.11 tools/gen_sample_images.py, access = sample_images.py.
gradient / blobs / shapes / checker_noisy — generated deterministically by tools/gen_sample_images.py.
Contains no third-party rights whatsoever.
skimage.data (BSD-3-Clause project · each image as below)Canonical source = https://github.com/scikit-image/scikit-image (skimage/data/; the license of each image is in
skimage/data/README.txt / LICENSE.txt) / API = https://scikit-image.org/docs/stable/api/skimage.data.html
| name | content | license / provenance |
|---|---|---|
coins |
Photo of Greek coins (a staple for blob/segmentation) | Bundled with scikit-image · effectively public domain (skimage/data/README.txt) |
camera |
“cameraman” (classic test image) | CC0 (photographer Lav Varshney). The CC0 version substituted in v0.18 out of copyright consideration |
page |
A scanned document page (binarization / OCR preprocessing) | Bundled with scikit-image · effectively public domain |
cell |
Quantitative phase imaging (derived from a digital hologram) | CC0 (public domain). Credit = Paul Müller, Mirjam Schürmann, Salvatore Girardo, Gheorghe Cojoc, Jochen Guck. Source paper = accurate assessment of the size/refractive index of spherical objects (quantitative phase imaging). Acquisition library = qpformat |
honest: The original photographers of
coins/pageare bundled by skimage as “public domain / no known copyright.” The strict tracing of the original source treatsskimage/data/README.txtas canonical (it may be updated across versions).
There are many classic image-processing benchmark images in the following. If you import any, verify each license individually (many are restricted to research/academic use). Fullseye does not currently bundle these (kept as reference only).
| dataset / repo | content | source paper / URL · license |
|---|---|---|
| scikit-image data | The classic set above + astronaut/coffee/chelsea, etc. | github.com/scikit-image/scikit-image (BSD-3; each image CC0/PD) |
| OpenCV samples | lena replacement · fruits · building, etc. | github.com/opencv/opencv samples/data/ (Apache-2.0) |
| USC-SIPI Image Database | Standard test images such as Baboon (Mandrill)/Peppers/cameraman | sipi.usc.edu/database (research use). Note: Lena is discouraged out of ethical consideration |
| BSDS500 (Berkeley Segmentation) | 500 natural images + human segmentations | Arbeláez, Maire, Fowlkes, Malik, “Contour Detection and Hierarchical Image Segmentation”, IEEE TPAMI 2011 (academic) |
| Set5 / Set14 / BSD68 / DIV2K | Super-resolution / denoising benchmarks | The respective SR/denoising papers (DIV2K = Agustsson & Timofte, CVPRW 2017) |
| MVTec AD (anomaly detection) | Industrial defect images | Bergmann et al., “MVTec AD”, CVPR 2019 (research-only · non-commercial, license check required) |
★ Caution: The sample images bundled with HALCON (MVTec) are proprietary, so they are not collected into Fullseye. The source papers for the ops / algorithms are already recorded in each backend’s docstring and in
docs/REFERENCES.md(RANSAC = Fischler & Bolles 1981, SGM = Hirschmüller, PPF = Drost 2010, etc.).