Provenance and licensing for everything in example_inputs/.

Two invariants this file records, both checkable:
  1. every file in example_inputs/ is LOADED by at least one workflow (as a
     widget value in workflows/ or subgraphs/) - unused assets are deleted or
     moved out, they are not parked here;
  2. every file in example_inputs/ has an entry below (a "a.png .. d.png"
     range header covers the whole series).

ONE FILE IS NON-COMMERCIAL: kitti_2011_09_26_drive_0009_poses.txt is
CC BY-NC-SA 3.0, inherited from KITTI. Everything else here is CC0 / CC BY /
CC BY-SA / public domain / synthetic. The repo's GPL-3.0 grant does not extend
to that file - see the License and provenance section of README.md.

Entries marked NOT SHIPPED are kept for provenance only: the asset lives in
tests/data/ (git-ignores it - see .gitignore) or was dropped entirely, and the
tests that want it skip when it is absent.

Model files (.onnx, cascades, LLM/VLM folders, checkpoints) are NOT in this
repo and are NOT part of example_inputs - their provenance and licenses live
in model_sources.txt at the repo root.

clock_wall_red.jpg   (used by: 08_line_segments_lsd, exercise_clock_reading)
  Title:   Modern functual kept battery driven wall clock
  Creator: Islander61
  License: CC BY-SA 4.0 (attribution + share-alike)
           https://creativecommons.org/licenses/by-sa/4.0/
  Source:  https://commons.wikimedia.org/wiki/File:Modern_functual_kept_battery_driven_wall_clock.jpg
  File:    https://upload.wikimedia.org/wikipedia/commons/c/c1/Modern_functual_kept_battery_driven_wall_clock.jpg
  Size:    1920x1279 (downscaled from original 5089x3391)

seamless_birds_sky.jpg   (used by: 17_seamless_clone - source patch)
  Title:   Birds flying at Sky 2.JPG
  Creator: Augustus Binu
  License: CC BY-SA 3.0 (attribution + share-alike)
           https://creativecommons.org/licenses/by-sa/3.0/
  Source:  https://commons.wikimedia.org/wiki/File:Birds_flying_at_Sky_2.JPG
  File:    https://upload.wikimedia.org/wikipedia/commons/2/28/Birds_flying_at_Sky_2.JPG
  Size:    540x294 (sky-only crop, downscaled from original 3604x2158;
           tree branch cropped out so the source border is uniform sky
           and seamless clone needs no mask)

seamless_sky_landscape.jpg   (used by: 17_seamless_clone - destination;
                         also 20_color_range_playground, exercise_clean_plate_matting)
  Title:   Clouds in the sky landscape.jpg
  Creator: Aammell
  License: CC0 1.0 (public domain dedication)
           https://creativecommons.org/publicdomain/zero/1.0/
  Source:  https://commons.wikimedia.org/wiki/File:Clouds_in_the_sky_landscape.jpg
  File:    https://upload.wikimedia.org/wikipedia/commons/f/f9/Clouds_in_the_sky_landscape.jpg
  Size:    1200x900 (downscaled from original 4000x3000)

water_coins.jpg   (used by: 22_watershed_playground, 26_hfs_segmentation)
  Title:   water_coins.jpg
  Creator: OpenCV project (sample data)
  License: Apache-2.0 (OpenCV) - sample data redistributed with the library
           https://www.apache.org/licenses/LICENSE-2.0
  Source:  https://github.com/opencv/opencv (samples/data/water_coins.jpg)
  Size:    252x312 (the canonical OpenCV watershed tutorial image)

inpaint_mask_holes.png   (used by: 18_inpainting_playground,
                         23_connected_components_list_vs_batch, 24_kmeans_clusters,
                         71_detection_postprocessing, 84_measure_and_export)
  Title:   inpaint_mask_holes.png
  Creator: generated (this repo) - synthetic mask image
  License: no third-party rights (synthetic binary mask; white spots on black)
  Source:  synthetic; bundled as example_inputs/inpaint_mask_holes.png
  Size:    768x768

text_page_shadow.jpg   (used by: 04_threshold_playground)
  Title:   Sunlight Illuminating Passage in Textbook (photo id 16196048)
  Creator: Aysenaz Bilgin
  License: Pexels License (free to use, no attribution required, modification
           allowed)
           https://www.pexels.com/license/
  Source:  https://www.pexels.com/photo/sunlight-illuminating-passage-in-textbook-16196048/
  File:    https://images.pexels.com/photos/16196048/pexels-photo-16196048.jpeg
  Size:    900x986 (cropped to the page - rows 22%..95% of the 4160x6240
           original - then downscaled)
  Why:     A page of dense printed text under a hard light shaft: the paper runs
           from ~14 to ~251 grey across one frame, so NO single global threshold
           separates ink from paper everywhere (Otsu blacks out ~82% of it).
           Adaptive mean/gaussian survive the ramp but lose the deepest shade,
           where the ink/paper gap falls below their fixed C - which is exactly
           what cv2.ximgproc.niBlackThreshold's contrast-scaled offset
           (mean + k*stddev) recovers. Niblack gets the text back everywhere and
           speckles the flat paper; Sauvola's r-normalisation does not. That
           four-way split is the workflow.
  Used in: workflows/04_threshold_playground.json (copy into ComfyUI/input/);
           tests/run_workflows.py::test_threshold_playground injects a synthetic
           ramp-plus-bars stand-in rather than reading this file

pexels-artful-homes-921250275-31663643.jpg   (used by: exercise_document_perspective)
  Title:   Lease agreement document with pen (photo id 31663643)
  Creator: Artful Homes
  License: Pexels License (free to use, attribution not required)
           https://www.pexels.com/license/
  Source:  https://www.pexels.com/@artful-homes-921250275/
  Size:    original download (used full-size)

calib_board_00.png .. calib_board_06.png   (used by: 49_chessboard_playground,
                         50_camera_calibration_roundtrip, 57_camera_pose_3d_preview,
                         exercise_ar_planar_cube)
  Title:   synthetic chessboard views with baked-in lens distortion
  Creator: generated (this repo)
  License: no third-party rights (synthetic renders)
  Source:  synthetic
  Size:    800x600 each
  Note:    the generator renders EIGHT views (N_VIEWS = 8); no workflow loads the
           eighth (calib_board_07.png), so it is not shipped here. Delete it again
           after re-running gen_calib_boards.py.

calib_board_07_extra_distorted.png   (used by: 49_chessboard_playground - the
                         deliberately over-distorted board)
  Title:   calib_board_07_extra_distorted.png
  Creator: generated (this repo) - the heavily distorted variant
  License: no third-party rights (synthetic render)
  Source:  synthetic
  Size:    800x600 (a calib_board_07 variant with extra distortion for chessboard playground)

demo_camera_params.json   (used by: 50_camera_calibration_roundtrip - the seed
                         the workflow's 'CV Load Camera Params (JSON)' reads on
                         the very first run, before its Save step has written
                         anything; not an image)
  Title:   demo_camera_params.json
  Creator: generated (this repo)
  License: no third-party rights (synthetic)
  Source:  synthetic; what cv2.calibrateCamera recovers from the eight
           calib_board_*.png views above (fx,fy = 802.1, 802.3 against a
           ground truth of 800, 800; cx,cy = 398.3, 299.9 against 400, 300;
           RMS 0.092 px).
  Size:    image_width 800, image_height 600 - it also makes the workflow's
           LoadCameraParams combo option exist, so validate_workflows passes on
           a fresh checkout.

circle_grid_asymmetric_4x11.png   (used by: 52_circle_grid_calibration - the
                         flat, orthogonal detection demo)
  Title:   synthetic ASYMMETRIC circle grid, 4x11
  Creator: generated (this repo)
  License: no third-party rights (synthetic render)
  Source:  synthetic; the 4x11 asymmetric pattern cv2.findCirclesGrid expects,
           drawn head-on with no perspective and no distortion, so the detector
           demo has a target whose centre order is unambiguous.
  Size:    640x560

circle_calib_00.png .. circle_calib_05.png   (used by: 52_circle_grid_calibration
                         - the MONO calibration branch)
  Title:   synthetic posed + distorted 4x11 circle-grid views
  Creator: generated (this repo)
  License: no third-party rights (synthetic renders)
  Source:  synthetic; the same 4x11 asymmetric board (32 mm spacing) projected
           through a KNOWN camera (fx = fy = 800, cx,cy = 320,240, barrel
           k1,k2 = -0.18, 0.08) from six different poses with cv2.projectPoints,
           so cv2.calibrateCamera has a ground truth to be scored against.
  Size:    640x480 each

circle_calib_L_00.png .. circle_calib_L_05.png +
circle_calib_R_00.png .. circle_calib_R_05.png
                     (used by: 52_circle_grid_calibration - the STEREO branch)
  Title:   synthetic stereo pairs of the same six circle-grid poses
  Creator: generated (this repo)
  License: no third-party rights (synthetic renders)
  Source:  synthetic; the six poses above shot twice through the same known
           camera, 100 mm apart on x, so cv2.stereoCalibrate recovers a
           baseline it can be checked against.
  Size:    640x480 each

omnidir_board_00.png .. omnidir_board_07.png   (used by: 54_omnidir_calibration)
  Title:   synthetic CATADIOPTRIC chessboard views (CMei omnidirectional model)
  Creator: generated (this repo)
  License: no third-party rights (synthetic renders)
  Source:  synthetic. Rendered through a known omnidirectional camera
           (fx = fy = 300, cx,cy = 400,300, xi = 0.8,
           k1,k2,p1,p2 = -0.12, 0.04, 0.001, -0.0008) by inverting CMei's model
           IN CLOSED FORM - deliberately not with cv2.omnidir.undistortPoints,
           which normalises to z = 1 and is up to 737 px wrong at the frame edge
           where rays pass the horizon. The generator asserts its inverse
           round-trips through cv2.omnidir.projectPoints to 0.0000 px and that
           the detected corners land within 1 px of it. Not taken from
           opencv_extra's testdata/cv/cameracalibration/omnidir: that repository
           ships no LICENSE file.
  Size:    800x600 each

omnidir_scene_L.png + omnidir_scene_R.png   (used by: 54_omnidir_calibration -
                         the omnidir stereo pair)
  Title:   synthetic omnidirectional stereo pair of a textured room
  Creator: generated (this repo)
  License: no third-party rights (synthetic renders)
  Source:  synthetic; the same known camera, twice, 60 units apart and slightly
           toed in, looking at a procedurally textured six-wall room with a
           nearer poster plane. Ray-traced per pixel (nearest positive plane
           intersection), so the pair has true parallax at a known baseline.
  Size:    800x600 each

omnidir_camera_params.json   (NOT SHIPPED - ground truth for
                          54_omnidir_calibration, read by the tests and never
                          loaded by the workflow; not an image)
  Title:   omnidir_camera_params.json
  Creator: generated (this repo)
  License: no third-party rights (synthetic)
  Source:  synthetic; K / xi / D / stereo R / stereo T the assets were rendered
           with, what cv2.omnidir.calibrate recovers from them, and where
           cv2.omnidir.projectPoints puts every board's inner corners (the
           tests compare their own detections against that instead of
           duplicating the scene geometry).
  Location: tests/data/omnidir_camera_params.json - it is regenerated with the
           boards, so it is kept OUT of git; example_inputs no longer ships it
           and the tests that read it (run_workflows.py::test_omnidir_calibration,
           smoke_test.py::test_cg_omnidir_calibrate) skip when it is absent.

fisheye_board_00.png .. fisheye_board_07.png   (used by: 53_fisheye_calibration)
  Title:   synthetic WIDE-ANGLE chessboard views (equidistant fisheye model)
  Creator: generated (this repo)
  License: no third-party rights (synthetic renders)
  Source:  synthetic. Each destination pixel is un-projected with
           cv2.fisheye.undistortPoints through a KNOWN camera (fx = fy = 320,
           cx,cy = 400,300, k1..k4 = -0.25, 0.05, -0.0125, 0.0025) and the flat
           board is sampled there, so the views obey the fisheye model exactly
           and cv2.fisheye.calibrate has a ground truth to be checked against
           (the generator asserts the rendered corners land within 1 px of
           cv2.fisheye.projectPoints). Deliberately NOT taken from
           opencv_extra's testdata/cv/cameracalibration/fisheye: that repository
           ships no LICENSE file and the folder is named after a third-party
           toolbox, so its images are not safe to redistribute here.
  Size:    800x600 each

fisheye_camera_params.json   (NOT SHIPPED - ground truth for
                         53_fisheye_calibration, read by the tests and never
                         loaded by the workflow; not an image)
  Title:   fisheye_camera_params.json
  Creator: generated (this repo)
  License: no third-party rights (synthetic)
  Source:  synthetic; the K/D the boards were rendered with, plus what the
           fisheye and pinhole calibrators recover from them. tests/run_workflows
           .py reads it so the assertions cite the generator, not a copy.
  Location: tests/data/fisheye_camera_params.json - it is regenerated with the
           boards, so it is kept OUT of git; example_inputs no longer ships it
           and the tests that read it (run_workflows.py::test_fisheye_calibration,
           smoke_test.py::test_cg_fisheye_calibrate) skip when it is absent.

box.png + box_in_scene.png   (used by: 32_shape_match_warp,
                         34_corner_detection_playground, 36_feature_matching_playground,
                         68_east_text_detection, 75_colorization, exercise_homography_ar,
                         exercise_homography_ransac - replaces copyrighted mobydick pair)
  Title:   box.png / box_in_scene.png
  Creator: OpenCV project (sample data)
  License: Apache-2.0 (OpenCV) - sample data redistributed with the library
           https://www.apache.org/licenses/LICENSE-2.0
  Source:  https://github.com/opencv/opencv (samples/data/)
  Size:    box.png 324x223, box_in_scene.png 512x384
           (the canonical OpenCV feature-matching / homography pair)

stereo_bethlehem_L.jpg + stereo_bethlehem_R.jpg   (used by: exercise_epipolar_geometry,
                         exercise_two_view_sfm)
  Title:   Street scene - Bethlehem (stereograph)
  Creator: unknown photographer; held by the Library of Congress
  License: Public domain (LoC "no known restrictions on publication")
  Source:  https://commons.wikimedia.org/wiki/File:Street_scene_-_Bethlehem_LCCN2004672895.tif
           (LCCN 2004672895)
  Size:    740x949 each - the left/right halves of the stereo card, cropped from
           the ~3840x2421 render (card border + centre seam removed). The two lenses
           give a genuine horizontal baseline (real parallax), replacing the earlier
           unverified "real-world-scene..." pair and the example.png placeholders.
           NOTE: the card halves were published pre-aligned, so the residual
           disparity is tiny (~+-3 px) - fine for F / epipolar geometry, but too
           flat for a disparity map; the stereo-rectification exercise now uses
           the synthetic stereo_gallery pair below instead.

stereo_gallery_L.png + stereo_gallery_R.png   (used by: 60_quasi_dense_stereo,
                         exercise_stereo_rectification)
  Title:   synthetic "gallery" stereo pair (this repo)
  Creator: generated (this repo)
  License: source paintings public domain (the art_ref_* set below); everything
           else (brick/plank/stripe textures, layout) is synthetic - no
           third-party rights
  Size:    960x540 each - textured planes at 4-12 m depth rendered through two
           pinhole cameras with a 0.35 m horizontal baseline (fx = 700 px ->
           ground-truth disparity ~20-61 px); the right camera has a small
           roll + pitch so uncalibrated rectification has visible work to do.

panorama_A.png + panorama_B.png   (used by: 48_panorama_playground -
                         replaces unverified UT-tower pair)
  Title:   Neuschwanstein Castle (photochrom print)
  Creator: Photoglob Zurich / Detroit Publishing Co.; held by the Library of Congress
  License: Public domain (LoC Photochrom Print Collection, no known restrictions)
  Source:  https://commons.wikimedia.org/wiki/File:Neuschwanstein_Castle_LOC_print.jpg
  Size:    820-wide each - two overlapping crops of the one photochrom; the right
           view (panorama_B) is perspective-warped to emulate a camera rotated about
           its optical centre, so a single homography aligns the pair.

art_ref_starry.png / art_ref_wave.png / art_ref_pearl.png + art_posters_wall.png
        (used by: 08_line_segments_lsd, 14_chromatic_aberration,
         20_color_range_playground, 32_shape_match_warp, 33_delaunay_voronoi,
         45_annotate_quad_warp, 46_affine_shape_warp, 47_paste_into_scene,
         85_data_charts, 86_quality_metrics, exercise_color_transfer,
         exercise_multi_poster_detection, exercise_object_detection_homography,
         exercise_pyramid_blend, exercise_seam_pipeline
         - replace copyrighted movie posters)
  References - three public-domain paintings (all public domain via commons.wikimedia.org):
    art_ref_starry.png : Vincent van Gogh, "The Starry Night" (1889)
    art_ref_wave.png   : Katsushika Hokusai, "The Great Wave off Kanagawa" (c.1831)
    art_ref_pearl.png  : Johannes Vermeer, "Girl with a Pearl Earring" (c.1665)
  Scene - art_posters_wall.png : synthetic composite (this repo) of the three paintings
    warped onto a generated plaster wall. Source paintings public domain; the wall
    texture is synthetic (no third-party rights). 1500x1050.

pedestrian_area.mp4
        (40 kalman tracking, 76 RAFT optical flow, 77 RIFE frame interpolation,
         exercise background subtraction, exercise background subtraction latent,
         exercise efficient video remap, exercise optical flow interpolation
         (+ comparison); replaces the unverified crosswalk clip)
  Title:   "pedestrian area" (video codec test sequence)
  Creator: Taurus Media Technik, via Xiph.org derf test media
  License: CC0 1.0 (public domain dedication)
           https://creativecommons.org/publicdomain/zero/1.0/
  Source:  https://commons.wikimedia.org/wiki/File:Video_Codec_Test_pedestrian_area_1080p25.y4m.webm
           (original http://media.xiph.org/video/derf/)
  Size:    a 2s, 640x360 slice (50 frames @ 25fps). (The pedestrian_frameA/B.png stills -
           frames 0 and 2 of the slice - were dropped; nothing referenced them.)
           Workflow 40 tracks the RED BICYCLE crossing left to right: it is the one
           strongly-coloured object on a near-constant heading (~16 px/frame, frames
           21..28), and it is genuinely occluded by pedestrians from frame 26, which
           is what gives that workflow a real detector failure to coast through.
           25 fps and smooth motion are why the flow/interpolation workflows keep this
           clip rather than the 10 fps people_walking.mp4 below.

people_walking.mp4
        (28 bgsegm background subtraction)
  Title:   vtest.avi (OpenCV sample clip)
  Creator: OpenCV contributors, samples/data
  License: Apache License 2.0 (attribution)
           https://www.apache.org/licenses/LICENSE-2.0
  Source:  https://github.com/opencv/opencv/blob/4.x/samples/data/vtest.avi
           (raw: https://raw.githubusercontent.com/opencv/opencv/4.x/samples/data/vtest.avi)
  Size:    200 frames, 512x288 @ 10fps - the first 20s of the original, cropped to the
           walkway band (x=224, y=128, 512x288 out of 768x576). The full-length 795-frame
           original stays in the local-only tests/data/ corpus.
           WHY A CROP: over the whole 768x576 view only 2.1% of pixels move, below the
           0.05-0.2 band 'CV Background Subtract (Batch)' calls a healthy
           foreground_fraction, so the workflow would have contradicted its own tooltip.
           The crop reads 0.050 and cuts the IMAGE batch from 1.6 GB to 354 MB.
           WHY 200 FRAMES: CNT/GSOC/LSBP need hundreds of frames to converge. On this
           clip they read 16.0/12.5/16.0% over the first 50 frames and 7.9/9.3/13.3%
           over all 200, against a 5.0% temporal-median reference - so the workflow can
           show a FAIR comparison and still demonstrate the failure by trimming.

kitti_2011_09_26_drive_0009_poses.txt   (used by: exercise_visual_odometry_video)
  Title:   KITTI raw drive 2011_09_26_drive_0009 - camera-frame ground-truth poses
  Creator: Geiger, Lenz, Stiller & Urtasun (KITTI); derived here from the drive's
           own oxts log
  License: CC BY-NC-SA 3.0 - **NON-COMMERCIAL**, inherited from KITTI. This is the
           ONLY non-commercial file in example_inputs; everything else is CC0 /
           CC BY / public domain / synthetic. Attribution and share-alike apply.
           https://creativecommons.org/licenses/by-nc-sa/3.0/
  Source:  https://www.cvlibs.net/datasets/kitti/raw_data.php
           Cite: Geiger et al., "Vision meets Robotics: The KITTI Dataset"
           https://www.cvlibs.net/publications/Geiger2013IJRR.pdf
  Size:    447 lines, 86 KB. One pose per frame: the 12 numbers of a 3x4 row-major
           camera-to-world matrix, first frame at the origin.
  Note:    DERIVED, not downloaded. KITTI raw ships oxts/data/*.txt (GPS/IMU
           packets); poses.txt is the format of the separate KITTI *odometry*
           benchmark, which uses different sequences. This file was built from this
           drive's oxts log - lat/lon through a Mercator projection to local metres,
           R from roll/pitch/yaw, frame 0 as the origin - then composed through
           calib_imu_to_velo and calib_velo_to_cam from 2011_09_26_calib.zip so the
           track follows the CAMERA centre rather than the IMU. The ~0.8 m lever arm
           rotates during turns, so an uncorrected IMU track is not a rigid offset of
           the camera track and no trajectory alignment can absorb the difference.
           The matching 48 MB video is NOT shipped; the workflow's Note says how to
           encode it from the drive archive.

traffic_frame_a.png + traffic_frame_b.png   (used by: 39_optical_flow_playground,
                         exercise_optical_flow_motion_segmentation)
  Title:   frames from car-detection.mp4 (sample-videos)
  Creator: Intel IoT Devkit
  License: CC BY 4.0
           https://creativecommons.org/licenses/by/4.0/
  Source:  https://github.com/intel-iot-devkit/sample-videos
  Size:    768x432 - two consecutive frames of a static traffic camera (a crossing car).

vo_frame0.png .. vo_frame3.png   (used by: exercise_visual_odometry -
                         replaces the forward-dashcam pair)
  Title:   View from Train (side window, station approach)
  Creator: uploaded to Wikimedia Commons
  License: CC BY-SA 4.0 (attribution + share-alike)
           https://creativecommons.org/licenses/by-sa/4.0/
  Source:  https://commons.wikimedia.org/wiki/File:View_from_Train_111450.ogv
  Size:    720x404 - four consecutive frames ~0.13s apart (stride 4 @ 30fps, ~t=221s)
           from the 480p transcode. The camera looks SIDEWAYS out the window during a
           slow station approach, so a static, richly-textured scene (boundary wall,
           parked car, signal gantry, litter-strewn ground) translates laterally with
           strong, consistent parallax. This is the geometry the essential matrix wants
           (lateral translation, static scene) - it replaces the earlier forward-driving
           dashcam frames, whose focus-of-expansion / moving sky were near-degenerate and
           gave as few as 9 tracked corners and inconsistent per-step directions. The new
           set tracks 430+ corners with 360+ RANSAC inliers on every step.

ar_grid_clip.mp4   (used by: exercise_ar_planar_video)
                     + ar_video_grid_0.png .. ar_video_grid_3.png (NOT SHIPPED)
  Title:   chessboard.avi (handheld 9x6 calibration board clip)
  Creator: Matej Smid (smidm)
  License: MIT License, (c) 2015 Matej Smid (attribution)
           https://opensource.org/license/mit
  Source:  https://github.com/smidm/video2calibration (example_input/chessboard.avi)
  Size:    ar_grid_clip.mp4 is a 4-frame 640x360 clip (source 1280x720@25fps,
           frames 127/130/133/136, downscaled 0.5x) for the workflow's
           LoadVideo. The board is detected once on frame 0 and KLT-tracked
           into the rest.
  Location: the four stills of the same frames are the TEST seed - they let
           tests/run_workflows.py::test_ar_planar_video feed the graph without
           a video decoder - so they live in tests/data/ar_video_grid_{0..3}.png
           and are not shipped. The stills and the clip come from the same
           fetch of the upstream .avi; the test skips when the stills are
           absent.

NH_64355_Brigadier_General_Smedley_D._Butler,_USMC.jpg
                     (used by: 24_kmeans_clusters, 65_yunet_face_detection,
                      66_face_recognition)
  Title:   NH 64355 Brigadier General Smedley D. Butler, USMC
  Creator: U.S. Marine Corps / Naval History and Heritage Command
  License: Public domain (work of the U.S. federal government)
  Source:  https://commons.wikimedia.org/wiki/File:NH_64355_Brigadier_General_Smedley_D._Butler,_USMC.jpg
  Size:    500x751 (single frontal portrait - one clear face for the detector)

OK_Sign.jpg   (used by: 67_handpose_estimation, 74_batch_dnn_style_transfer)
  Title:   OK Sign
  Creator: Steven H. Keys (User:Steevven1)
  License: CC BY 2.5 (attribution)
           https://creativecommons.org/licenses/by/2.5/
  Source:  https://commons.wikimedia.org/wiki/File:OK_Sign.jpg
  Size:    416x411 (a single upright hand making the OK gesture - one clear palm
           for the MediaPipe palm detector, then 21-keypoint hand-pose regression)

market_scene.jpg + market_template.jpg + market_template_1p5x.jpg
        (used by: 32_shape_match_warp, 35_keypoint_detectors, 37_match_template,
         38_match_template_multiscale, 42_rapid_model_tracking,
         61_ppf_pose_estimation, exercise_histogram_tracking,
         exercise_image_registration_ecc - replace example.png)
  Title:   Fruits and vegetables at market
  Creator: Santeri Viinamaki
  License: CC BY-SA 4.0 (attribution + share-alike)
           https://creativecommons.org/licenses/by-sa/4.0/
  Source:  https://commons.wikimedia.org/wiki/File:Fruits_and_vegetables_at_market.jpg
  Size:    market_scene.jpg 1000x750 (downscaled from 4592x3448). The templates are the
           unique round red sale tag cropped from the scene: market_template.jpg at native
           size (37_match_template, best match score ~1.0),
           market_template_1p5x.jpg upscaled 1.5x
           (38_match_template_multiscale, found at best_scale ~0.67).

qr_code_wikipedia.png   (used by: 81_qr_code_detection)
  Title:   QR code for mobile English Wikipedia
  Creator: Brdall (Wikimedia Commons)
  License: Public domain (creator released it for any use, no conditions)
  Source:  https://commons.wikimedia.org/wiki/File:QR_code_for_mobile_English_Wikipedia.svg
  File:    https://upload.wikimedia.org/wikipedia/commons/thumb/d/d0/QR_code_for_mobile_English_Wikipedia.svg/250px-QR_code_for_mobile_English_Wikipedia.svg.png
  Size:    250x250 (smallest PNG render of the SVG; the render is black-on-
           transparent, so the alpha was flattened onto white to give an opaque
           QR the detector can read). Decodes to http://en.m.wikipedia.org.
  Models:  detect_2026april.onnx + sr_2026april.onnx - not in this repo,
           see model_sources.txt

walking_the_dog.jpg   (used by: 16_npr_filters_playground, 69_yolov4_detection,
                         70_yolo_seg, 79_vlm_caption)
  Title:   Walking the dog
  Creator: Helgi Halldorsson from Reykjavik, Iceland
  License: CC BY-SA 2.0 (attribution + share-alike)
           https://creativecommons.org/licenses/by-sa/2.0/
  Source:  https://commons.wikimedia.org/wiki/File:Walking_the_dog_(1945527533).jpg
  File:    https://upload.wikimedia.org/wikipedia/commons/thumb/2/24/Walking_the_dog_%281945527533%29.jpg/960px-Walking_the_dog_%281945527533%29.jpg
  Size:    960x1440 (960px thumbnail render, downscaled from original 2592x3888).
           A person walking a dog - one clear person + dog for the detector.
  Models:  yolov4.onnx / yolov4x-mish.onnx - not in this repo,
           see model_sources.txt

coco-labels-2014_2017.txt   (used by: 69_yolov4_detection, 70_yolo_seg and the
                             'CV YOLO Detection' / 'CV YOLO-Seg Inference'
                             blueprints - class-name list for CV Load Labels)
  Title:   COCO object-category labels (2014 + 2017 releases)
  Creator: Microsoft COCO dataset; label list maintained by amikelive
  License: MIT (this repository: https://github.com/amikelive/coco-labels)
           https://opensource.org/license/mit
  Source:  https://github.com/amikelive/coco-labels/tree/master (coco-labels-2014_2017.txt)
  Size:    80 lines - one class name per line (line N = class id N). The 2014 and
           2017 releases share the same 80 categories, so the repo merges them into
           one file; the COCO annotations themselves are Microsoft's (CC BY 4.0),
           this is just the plain class-name list it ships alongside.

dog_leaping.jpg   (used by: 72_dnn_deblur, 73_dnn_style_transfer)
  Title:   Dog Leaping
  Creator: Harry Walker (photographer, 1894-1971), 1920s
  License: Public domain (not copyrighted)
  Source:  https://commons.wikimedia.org/wiki/File:Dog_Leaping.jpg
  File:    https://upload.wikimedia.org/wikipedia/commons/thumb/e/e2/Dog_Leaping.jpg/960px-Dog_Leaping.jpg
  Size:    960x1645 (960px thumbnail render, downscaled from original 3168x5430).
           An old grainy film photo of a dog leaping for a ball - motion blur +
           film grain give the NAFNet restorer something real to clean up.
  Model:   deblurring_nafnet_2025may.onnx - not in this repo,
           see model_sources.txt

brick_wall_nor_gl_1k.jpg   (used by: 44_relative_remap)
  Source: Poly Haven (CC0)
  URL: https://polyhaven.com/a/brick_wall_001
  Download: https://dl.polyhaven.org/file/ph-assets/Textures/jpg/1k/brick_wall_001/brick_wall_001_nor_gl_1k.jpg
  Author: Dimitrios Savva, Rob Tuytel
  License: CC0 - https://creativecommons.org/publicdomain/zero/1.0/
  Used in: 44_relative_remap.json

aruco_board_5x7_DICT_6X6_250.jpg was generated in blender
  Author: bmad4ever
  License: CC0 - https://creativecommons.org/publicdomain/zero/1.0/
  Used in: 62_aruco_3d_render, 63_ar_model_annotation, exercise_aruco_board_cube

hdr_stlouis_ev-4.72.jpg / hdr_stlouis_ev-1.82.jpg /
hdr_stlouis_ev+1.51.jpg / hdr_stlouis_ev+4.09.jpg
                     (used by: 15_hdr_playground)
  Title:   St. Louis Arch, multi-exposure bracket (EV -4.72 / -1.82 / +1.51 / +4.09)
  Creator: Kevin McCoy (Wikimedia user Kmccoy)
  License: CC BY-SA 3.0 (attribution + share-alike)
           https://creativecommons.org/licenses/by-sa/3.0/
  Source:  https://commons.wikimedia.org/wiki/File:StLouisArchMultExpEV-4.72.JPG
           https://commons.wikimedia.org/wiki/File:StLouisArchMultExpEV-1.82.JPG
           https://commons.wikimedia.org/wiki/File:StLouisArchMultExpEV%2B1.51.JPG
           https://commons.wikimedia.org/wiki/File:StLouisArchMultExpEV%2B4.09.JPG
  Size:    1200x900 each (downscaled from the 2816x2112 originals)
           Four tripod exposures of the same scene - the canonical OpenCV HDR
           tutorial bracket. Stacked darkest-to-brightest into one batch; the
           EV offsets set the exposure_times (2^EV) fed to the Debevec merge.

gingham_check_diff_1k.jpg   (used by: 13_demosaicing_playground)
  Title:   Gingham Check (diffuse texture, 1k jpg)
  Creator: Poly Haven (Dario Barresi et al.)
  License: CC0 - https://creativecommons.org/publicdomain/zero/1.0/
  Source:  https://polyhaven.com/a/gingham_check
  Download: https://dl.polyhaven.org/file/ph-assets/Textures/jpg/1k/gingham_check/gingham_check_diff_1k.jpg
  Size:    1024x1071. A regular green/white woven check pattern - deliberately a
           demosaicing WORST CASE: fine periodic high-frequency detail is exactly
           what aliases into zipper/moire colour fringing, so bilinear vs VNG vs
           edge-aware demosaicing visibly separate. Mosaiced in-graph into a
           synthetic RGGB Bayer CFA (no licence-safe real Bayer sample exists),
           then reconstructed and scored (PSNR/absdiff) against itself.

MilkTruck.glb        (used by: 42_rapid_model_tracking, 57_camera_pose_3d_preview,
                      61_ppf_pose_estimation, 62_aruco_3d_render,
                      63_ar_model_annotation, 64_ar_depth_occlusion,
                      exercise_homography_ar; copy into ComfyUI/input/3d/)
  Title:   Milk Truck - de-branded derivative of the glTF-Binary sample model
           "Cesium Milk Truck"
  Creator: Cesium (2017); logo removed 2026
  License: CC BY 4.0 International - attribution required.
           https://creativecommons.org/licenses/by/4.0/
           The UPSTREAM file adds a trademark limitation on the Cesium logo
           texture (LicenseRef-LegalMark-Cesium), which is NOT covered by
           CC BY and would need Cesium's permission to redistribute. This copy
           removes it instead: the four logo roundels are painted out of the
           single baked 2048x2048 JPEG with the surrounding body colour, and
           the branded identifiers are renamed (glTF node + mesh
           Cesium_Milk_Truck -> Milk_Truck, image CesiumMilkTruck.jpg ->
           MilkTruck.jpg), so only the CC BY 4.0 terms apply. Geometry, UVs
           and materials are untouched - 4823 vertices / 3624 triangles, bbox
           diagonal 6.1784, byte-identical accessors - so every measurement
           taken against the original still holds.
  Source:  https://github.com/KhronosGroup/glTF-Sample-Assets (Models/CesiumMilkTruck)
  File:    https://raw.githubusercontent.com/KhronosGroup/glTF-Sample-Assets/main/Models/CesiumMilkTruck/glTF-Binary/CesiumMilkTruck.glb
  Size:    415 KB; mesh bounds ~4.9 x 2.8 x 3.0 scene units - the same order of
           magnitude as the 9x6 calibration board with spacing 1.0 (8 x 5 units),
           so the solvePnP camera distance (~10-17 units) frames it sensibly.
           The 3D loaders only list files under ComfyUI/input/3d/ - keep the
           copy there in sync with this one.

stereo_scene_L_000000.png + stereo_scene_R_000000.png + stereo_scene_L_000001.png + stereo_scene_R_000001.png
        + left_images.csv + right_images.csv
        (used by: 58_stereo_bm_vs_sgbm, 59_disparity_refinement,
         64_ar_depth_occlusion, exercise_stereo_multiview,
         exercise_stereo_pointcloud)
  Title:   StereoGeo-CARLA (stereo driving scenes from the CARLA simulator)
  Creator: Meddourimane et al.
  License: CC BY 4.0
           https://creativecommons.org/licenses/by/4.0/
  Source:  https://huggingface.co/datasets/meddourimane/StereoGeo-CARLA
  ArXiv:   https://arxiv.org/abs/2606.14619
  Size:    640x480 each - synthetic driving scenes rendered in the CARLA simulator.
           The full dataset has 12,220 stereo pairs; two pairs (000000.png, 000001.png)
           were extracted from batch_1.zip as examples. CSV files contain per-image
           calibration: intrinsics from vertical FOV (pinhole model, no distortion)
           and extrinsics (R, T) from the right camera's relative pose.

nyu_depth_rgb.jpg + nyu_depth_map.png   (used by: exercise_monocular_depth_3d)
  Title:   NYU Depth V2 (indoor RGB-D scene, sample 10 - hallway/room with
            objects at varying distances: 1-9.5 m depth range)
  Creator: Nathan Silberman et al. (NYU), redistributed via jagennath-hari/nyuv2
  License: MIT (NYU Depth V2 preprocessed dataset)
           https://opensource.org/license/mit
  Source:  https://huggingface.co/datasets/jagennath-hari/nyuv2
           (train split, sample 10)
  Camera:  https://huggingface.co/datasets/jagennath-hari/nyuv2/blob/main/camera_params.json
           fx=518.86, fy=519.47, cx=325.58, cy=253.74 (640x480 pinhole)
  Size:    640x480 each. The depth map is uint16 in millimeters (range
           ~1-9.5 m) from a Kinect sensor. The scene has objects at multiple
           distances giving a clearly 3D point cloud via DepthTo3D.

synthetic_color_checker.png   (NOT SHIPPED - no workflow loads it any more.
                         Kept here because it is an ingredient of
                         color_cast_posters.png below, which does ship)
  Title:   Synthetic Macbeth Color Checker (6x4 grid)
  Creator: generated (this repo) - Python/OpenCV script
  License: no third-party rights (synthetic color patches)
  Source:  synthetic; sRGB values from ColorChecker Wikipedia article;
           black border added for cv2.mcc.CCheckerDetector compatibility
  Size:    760x560 (6 patches x 100px + margins + border, 4 rows x 100px + margins + border)

color_cast_posters.png   (used by: 11_color_correction_basic,
                         12_white_balance_and_tone - the realistic demo)
  Title:   Art posters wall with synthetic ColorChecker + tungsten color cast
  Creator: generated (this repo) - composite of art_posters_wall.png
           + synthetic ColorChecker chart, then tungsten BGR cast
           (B=0.8, G=0.9, R=1.1)
  License: no third-party rights (synthetic composite)
  Source:  synthetic; posters from art_posters_wall.png (see above),
           chart from synthetic_color_checker.png with white card border,
           scaled to 40% image width and centered bottom; global tungsten
           cast applied so the CCM correction is visibly warm -> neutral.
           cv2.mcc detects the chart at this scale with the cast.
  Size:    1500x1050 (same as art_posters_wall.png)

boat1.jpg .. boat6.jpg   (used by: 48_panorama_playground -
                         the 6-view boat panorama)
  Title:   boat1.jpg .. boat6.jpg (stitching test set)
  Creator: OpenCV project (opencv_extra test data)
  License: Apache-2.0 (OpenCV) - sample data redistributed with the library
           https://www.apache.org/licenses/LICENSE-2.0
  Source:  https://github.com/opencv/opencv_extra/tree/5.x/testdata/stitching
  Size:    ~700 KB each (3888x2592); six overlapping views of a boat/harbour
           scene, captured with horizontal camera rotation. Sequential overlap
           is strong enough for pairwise feature matching + RANSAC homography.
  Why:     boat1 + boat2 overlap by ~2/3 of the frame AND the camera re-metered
           between the two shots: aligned, the shared scene is ~14% darker and
           slightly cooler in boat2 (a per-channel lift of ~ +13/+17/+20 B/G/R).
           A real auto-exposure/white-balance drift, not a simulated one - so
           the pair is the honest input for any exposure-matching measurement
           (see decisions.md 2026-09-08 for the numbers it produced).
  Used in: exercise_multi_image_stitching.json (foreach-loop multi-stitch);
           48_panorama_playground.json (six-view auto stitch);
           exercise_stitch_two_views.json (boat1+boat2 alignment + union)

synthetic_radial.png + radial.yaml   (used by: 14_chromatic_aberration,
                         16_npr_filters_playground - real OpenCV test data)
  Title:   Synthetic radial chromatic aberration + calibration
  Creator: OpenCV project (opencv_extra test data)
  License: Apache-2.0 (OpenCV)
           https://www.apache.org/licenses/LICENSE-2.0
  Source:  https://github.com/opencv/opencv_extra/tree/5.x/testdata/cv/cameracalibration/chromatic_aberration
  File:    radial.yaml  - degree-11 polynomial calibration (78 coefficients per
           channel), from real camera calibration. image_width=600, image_height=423.
           synthetic_radial.png - synthetic image with known radial CA matching
           the calibration.
  Size:    600x423 each
  Used in: 14_chromatic_aberration.json (row 2: real-data round-trip test)

barcode_book.jpg + barcode_bottle_1.jpg
                     (used by: 83_barcode_decode - real OpenCV test data with
                      GROUND TRUTH)
                     + barcode_result.csv (NOT SHIPPED - the expected-output
                      table, read by the tests, never loaded by the graph)
  Title:   barcode single-symbol test set
  Creator: OpenCV project (opencv_extra test data)
  License: Apache-2.0 (OpenCV) - sample data redistributed with the library
           https://www.apache.org/licenses/LICENSE-2.0
  Source:  https://github.com/opencv/opencv_extra/tree/5.x/testdata/cv/barcode/single
  Files:   barcode_book.jpg   - EAN-13 on a book back cover, 640x480
           barcode_bottle_1.jpg - EAN-13 on a curved bottle label, 800x600
           barcode_result.csv - the upstream ground-truth digits, kept so the
           tests can assert the EXACT decoded string instead of "something was
           found" (book = 9787115279460, bottle_1 = 6922255451427). It is
           test-only data, so it lives in tests/data/barcode_result.csv and is
           not shipped in example_inputs.
  Used in: 83_barcode_decode.json; tests/smoke_test.py::test_cg_barcode_detect
           and tests/run_workflows.py::test_barcode_decode assert the digits.

faces_butler_lejeune.jpg   (used by: 66_face_recognition - a MULTI-PERSON photo
                         containing the man in NH_64355_...Butler...jpg and others)
  Title:   Gen. Smedley Butler & Gen. Lejeune, 12-28-25
  Creator: National Photo Company Collection (Library of Congress)
  License: Public domain (LCCN 2016841610, no known restrictions)
  Source:  https://commons.wikimedia.org/wiki/File:Gen._Smedley_Butler_%26_Gen._Lejeune,_(12-28-25)_LCCN2016841610.jpg
  Size:    downscaled from 5259x4209 to 1200x960, JPEG quality 92
  Why:     paired with the existing Butler portrait it makes a self-contained
           recognition test: YuNet finds 5 faces, SFace identifies EXACTLY ONE
           of them as the gallery person (cosine 0.65 against 0.02-0.17 for the
           rest), so the assertion is decisive rather than "a match happened".
  Used in: 66_face_recognition.json; tests/smoke_test.py::test_cg_sface_recognition

matte_shot_clean.png / matte_plate_clean.png / matte_shot_a.png /
matte_plate_a.png / matte_alpha_gt.png
                     + matte_shot_b.png / matte_plate_b.png (NOT SHIPPED)
                     (used by: exercise_clean_plate_matting - single-pair matting;
                      matte_plate_a.png is ALSO 10_ximgproc_edge_aware_filters'
                      input - its plate noise beside the picture frames' hard
                      edges is what separates an edge-aware smoother from a blur)
  Title:   Clean plate matting scene (takes, clean plates, ground-truth alpha)
  Creator: generated (this repo)
  License: no third-party rights beyond the two backdrops it composites
  Source:  synthetic. Backdrop A is art_posters_wall.png, backdrop B is
           market_scene.jpg (both above), each cropped to 512x384.
           The subject is drawn at 4x and area-downsampled so its edges carry
           REAL fractional coverage, plus a 0.42-alpha glass panel and 26 hair
           wisps - the cases a threshold-and-feather mask cannot express, and
           the panel is UNIFORMLY translucent, which only the backdrop's own
           texture (showing through it) can resolve from a single pair.
           *_clean.png is the ideal pair: a noise-free take and the literal
           background pixels behind it, so the difference is exactly zero
           outside the subject. matte_shot_a/b.png carry sensor noise and
           matte_plate_a/b.png are the same backdrops shot with the camera
           nudged (2-3 px, <0.5 deg, a touch of scale and perspective) and the
           lamps drifted (a different per-channel gain AND bias each), which is
           exactly what the workflow has to undo. The B pair exists only for
           the two-background 'CV Triangulation Matte' blueprint, which no
           shipped workflow instantiates, so matte_shot_b.png and
           matte_plate_b.png live in tests/data/ - the generator writes the A
           pair to example_inputs and the B pair there. The test that wants
           them skips when they are absent.
  Size:    512x384 each; matte_alpha_gt.png is the 8-bit alpha the scene was
           built from, used only to score the result.
  Why:     the ground truth makes the exercise measurable rather than
           plausible-looking: mean |alpha - truth| 0.0035 on the ideal pair
           (and exactly 0 at every byte-identical pixel), 0.148 raw / 0.0059
           aligned + cleaned on the nudged one.
  Used in: exercise_clean_plate_matting, 10_ximgproc_edge_aware_filters
           (matte_plate_a.png only);
           tests/run_workflows.py::test_clean_plate_matting_ideal_pair,
           ::test_clean_plate_matting_real_pair,
           ::test_triangulation_matte_blueprint

gst_input.jpg + gst_result.jpg   (used by: 27_anisotropic_segmentation -
                         gradient structure tensor. gst_result.jpg is the published
                         reference figure and is NOT shipped; the test that scores
                         against it skips when it is absent)
  Title:   Anisotropic image segmentation tutorial input + published result figure
  Creator: OpenCV project (Karpushin Vladislav) - tutorial illustrations
  License: Apache-2.0 (OpenCV) - doc images redistributed with the library
           https://www.apache.org/licenses/LICENSE-2.0
  Source:  https://github.com/opencv/opencv
           (doc/tutorials/imgproc/anisotropic_image_segmentation/images/)
  Size:    both 320x320
  Note:    gst_result.jpg is the figure PUBLISHED with the tutorial, kept as
           ground truth: only the C++ sample's inRange band reproduces it
           (corr 0.997). The python sample passes HighThr as cv.threshold's
           maxval, never bounds the angle, and does not match (corr 0.785).

text_page_sharp.jpg + text_page_defocus.jpg + text_page_motion.jpg
                     (used by: 19_fourier_playground,
                      exercise_match_two_views_over_a_mask, and the in-flight
                      wiener-filter / motion-deblur test workflows)
  Title:   deconvolution test crop, sharp / disk-blurred / motion-blurred
  Creator: generated (this repo), from text_page_shadow.jpg above
  License: derived from text_page_shadow.jpg (Pexels License - modification
           allowed)
           https://www.pexels.com/license/
  Source:  synthetic. The SAME crop (rows 60..776, cols 180..814) three ways:
           _sharp is the untouched crop and the ground truth for both filters,
           _defocus is it convolved with a DISK of radius 6, _motion with a
           LINE of LEN=21 px at THETA=15 degrees (plus sigma 0.5 noise and a
           JPEG q92 round trip).
  Size:    634x716
  Used in: wiener_filter_test.json / motion_deblur_test.json (deconvolution);
           exercise_match_two_views_over_a_mask.json uses _sharp as the TILED-
           processing subject - a printed page is the classic tile-enhanced
           image, and its two halves hold genuinely different ink/paper
           distributions, so a per-tile white balance really does diverge
           (per-channel gains ~0.39/0.57/0.58 vs ~1.03/1.01/1.01)
  Note:    the blur parameters are CONSTANTS in those two generators, so the
           radius / LEN / THETA the workflows type are known-good rather than
           guessed - which matters because a PSF size is a LENGTH and does not
           survive a change of resolution (the OpenCV tutorials quote R=53 and
           LEN=125/78 for photos they do not ship). A forward fit over the
           shipped _motion asset bottoms out at LEN=21 THETA=15, RMSE 1.1/255.

period_input.jpg + period_result.jpg
                     (used by: 19_fourier_playground and the in-flight
                      periodic-noise test workflow. period_result.jpg is the
                      published reference figure and is NOT shipped)
  Title:   Periodic noise removing filter tutorial input + published result
  Creator: OpenCV project (Karpushin Vladislav) - tutorial illustrations
  License: Apache-2.0 (OpenCV) - doc images redistributed with the library
           https://www.apache.org/licenses/LICENSE-2.0
  Source:  https://github.com/opencv/opencv
           (doc/tutorials/imgproc/periodic_noise_removing_filter/images/,
           period_input.jpg and period_output.jpg - renamed _result here to
           match the gst_result.jpg convention above)
  File:    https://github.com/opencv/opencv/blob/23212f1427ccb1aa027d91e0737f08b1af07d44c/doc/tutorials/imgproc/periodic_noise_removing_filter/images/period_input.jpg
  Size:    both 640x471 (cropped to 640x470 by the tutorial's `rows & -2`)
  Note:    a halftone-screened newspaper photo - the printing screen is the
           periodic noise. period_result.jpg is the figure PUBLISHED with the
           tutorial and is kept as ground truth: the shipped workflow reaches
           0.971 correlation with it, against 0.737 for the input.
  Trap:    the tutorial's own notch coordinates - (705, 458), (850, 391),
           (993, 325) - do NOT belong to this file. All three are past the
           right edge of its 640-wide spectrum; they were measured on a larger
           photo the tutorial does not ship, exactly like the R=53 disk radius
           in the out-of-focus deblur tutorial. The workflow therefore does
           not type any coordinate: 'CV Detect Spectral Peaks' reads them off
           the log PSD (outliers at mean + 3.5 std), and lands on the same
           three families the hand measurement found - (403, 174), (321, 113)
           and (155, 233), radius 21.
