fullseye

Fullseye Gallery

日本語 · English

Purpose of this page: every figure shown here is a real output that can be regenerated by running tools/gen_article_assets.py (article montages, hero copies, thumbnails) and tools/gen_showcase_gifs.py (turntable GIFs). There are no mockups, hand-drawn art, or synthetic composites whatsoever (honest disclosure discipline; see the docstring of each generating script for details).

The explanatory articles, to keep their loading cost down, use 720px-wide thumbnails (docs/articles/assets/thumbs/), while the full-size images and the “what each panel represents” explanations are consolidated on this page.

1. Article montages (full size + panel notes)

1.1 physical_ai_montage.png — Physical AI sensor simulation

Generated by: tools/gen_article_assets.py::build_physical_ai_montage(). Each module’s run_*_demo() is actually executed, rendering a MuJoCo scene and then processing it through a sensor model; the real outputs are arranged in a 2x3 grid.

physical_ai_montage

Panel Module / op Meaning of the displayed numbers
LiDAR — range image to 3D point cloud lidar_sim.py n_points = number of reconstructed 3D points, channels = number of scan channels (vertical resolution), hit_ratio = fraction of emitted rays that hit an object
Stereo depth — block matching stereo_sim.py depth_corr = correlation between block-matching estimated depth and ground truth, median_err_m = median of the depth error (m)
Event camera (DVS) — per-pixel change events event_camera.py n_events = total number of DVS events generated, edge_corr = correlation between event density and edge strength
Focus stacking — depth-from-focus focus_stack.py sharpness_gain = sharpness multiplier after focus stacking, depth_focus_corr = correlation between estimated depth and focus position
Polarization camera — DoLP / AoLP polar_cam.py mean_dolp = mean degree of linear polarization (0–1), stokes_roundtrip = Stokes-vector round-trip reconstruction accuracy (1.0 = perfect match)
Camera + IMU sensor fusion — Kalman filter sensor_fusion.py RMSE (cm) after Kalman fusion, compared with the RMSE of the position sensor alone (showing that fusion yields a smaller error)

1.2 vision_ops_montage.png — classic 2D vision op chain

Generated by: tools/gen_article_assets.py::build_vision_ops_montage(). A real op chain is applied via fullseye.apply() to the bundled sample image coins (skimage.data, BSD).

vision_ops_montage

Panel op Content
Input — sample image (input) coins.png as-is
Gaussian smoothing gaussian Gaussian smoothing (a=0.3, b=0.0)
Edge magnitude — Sobel sobel_amp Sobel edge strength on the smoothed image
Segmentation — Otsu threshold otsu Otsu thresholding on the smoothed image
Connected components scipy.ndimage.label (on the otsu regions) Number of connected components = number of coins detected
Sub-pixel contours + measurement edges_sub_pix → select_contours After sub-pixel contour extraction (a=0.2), a length threshold a=0.7 removes short contours arising from the engraved texture and keeps only the outer rim. Contour count and mean blob area (px) are shown

1.3 itokawa_montage.png — 3D ops on the real point cloud of asteroid 25143 Itokawa

Generated by: tools/gen_article_assets.py::build_itokawa_montage(). The data is studio_assets/sample_3d/itokawa_points.npy (a measured point cloud derived from the JAXA Hayabusa / Gaskell shape model, float32, 3000 points). The same computations as examples_3d/itokawa_*.py (curvature3d / match3d / metrics3d) are run directly on the spot, and the real numbers are burned into the captions.

itokawa_montage

Panel Module / op Meaning of the displayed numbers    
Itokawa — raw point cloud itokawa_points.npy The real point cloud shown as-is as a 3D scatter (colored by distance from the origin, with rock-like shading). Point count and bounding dimensions (m)    
Surface curvature curvature3d.curvedness Colored by the curvature strength (curvedness) of each point. mean/std are the spread of the distribution, neighbor coherence r is the curvature correlation with neighboring points (high for a real surface, ~0 for random data — a verification item of the corresponding examples_3d/itokawa_curvature.py)    
Self-registration (ICP) match3d.icp_point2point_3d A scan obtained by applying an unknown 30-degree rotation and sensor noise to the reference point cloud is aligned back with ICP (left = before, right = after). rot err = error of the recovered rotation from ground truth (degrees), RMSE = final residual (m)    
Canonical pose (PCA axes) match3d.moment_axes The principal inertia axes (red = longest axis / green / blue, length proportional to the square root of the eigenvalues) overlaid on the real point cloud. principal-axis ratio = eigenvalue ratio of the longest to the second-longest axis, axis recovery = the degree to which the principal axes could be recovered after applying an unknown 50-degree rotation ( cos , 1.0000 = perfect match)

1.4 op_taxonomy.png — op taxonomy map (treemap)

Generated by: tools/gen_article_assets.py::build_op_taxonomy(). ops.REGISTRY (2D) and ops3d.OPS3D (3D) are actually imported to tally the op count per category, and drawn as an area-proportional squarified treemap (no external library; the Bruls/Huizing/van Wijk 1999 algorithm implemented from scratch on top of matplotlib). 2D is the set obtained by deduplicating ops.REGISTRY by op name (the same rule as ops.RT; because four same-named ops are re-registered across categories in REGISTRY, the last one wins), and 3D counts ops3d.OPS3D (a dict, category field) as-is. The script verifies that the totals match the numbers in the article body (currently 2D 860 / 3D 310) (the article body is the single source of truth; changed from the hardcoded 731/265 on 2026-09-03) (a safeguard so that a mismatch reveals that either the registry or the article number is stale).

op_taxonomy

The left side is 2D (ops.py, blue tones), the right side is 3D (ops3d.py, orange tones). The area of a rectangle = the op count of that category, and the label is “category name + op count” (when a rectangle is too small, the label is omitted to avoid cramming in unreadable text). The largest categories in 2D are halcon_ext (81), region (76), and features (71). The largest in 3D are geometry (23), render (14), and transform (12).

1.5 halcon_coverage_chart.png — per-chapter bars of HALCON coverage

Generated by: tools/gen_article_assets.py::build_halcon_coverage_chart(). The original instruction assumed using the covered field of fullseye/data/halcon_graph.json, but reading it actually gives only 252/2313 (10.9%), which did not match the article’s measured value (982/2313 = 42.5%) (honest disclosure — the result of verifying against the real data instead of taking the instructed reference at face value). What actually generates docs/HALCON_COVERAGE.md is halcon_coverage.py (which matches the 2313 ops of the real scrape result in data/halcon_operators.json against Op.halcon), so that is re-run on the spot to obtain the true per-chapter covered/total, and after generating the graph it is also cross-checked with the “982 / 2313 (42.5%)” figure written in docs/HALCON_COVERAGE.md via assert (the log prints cross-check OK when the script runs).

halcon_coverage_chart

The horizontal bars are sorted by coverage rate (covered/total) descending, with the actual n/N noted at the right end. Of all 30 chapters, Regions (105/106), Morphology (42/44), and Filters (186/196) are at the top, while the 8 chapters System, Classification, OCR, Control, Tuple, File, Image Source, and Develop have 0 coverage (HALCON’s non-algorithmic chapters or areas imgevolve does not yet support). The numbers are the tally results as-is.

1.6 op_sampler_2d.png — 2D op output sampler (24 category representatives)

Generated by: tools/gen_article_assets.py::build_op_sampler_2d(). A representative op from 24 categories mechanically selected out of 46 categories is actually applied to the bundled sample image coins (skimage.data, BSD). The selection is not by appearance: the first-appearance order of categories is recorded in the registration order of ops.REGISTRY, and for each category the first op that passed fullseye.apply() without an exception is adopted (ops that do not run fall back to the next candidate within the same category).

op_sampler_2d

24 tiles: identity (misc) gaussian (smoothing) median (rank) gerode (morphology) sobel_mag (edges) gamma (gray) lowpass (frequency) std_filter (texture) threshold (segmentation) reg_erode (region) blob_count (features) edges_sub_pix (contour) ncc_locate (matching) rotate_img (geometry) classify_shape (classification) decode_barcode (barcode) vol_gaussian (3d) abs_image (arithmetic) add_noise_white (noise) cfa_to_rgb (color) xsk_inpaint (restoration) xcv_stylization (artistic) xmh_zernike (texture/shape-feature) xmh_pftas (texture-feature). image/region outputs are shown as images as-is, feature (scalar/vector) outputs have their numbers burned into the tile (e.g., blob_count=244 items, classify_shape=0.1044), and the contour output (edges_sub_pix) has the real XLD points burned into the overlay (the points themselves are shown without connecting them with lines, because connecting the scan-order points with lines produces false chords). vol_gaussian applies an op that is originally intended for a 3D volume directly to a single 2D image (equivalent to one slice), and its appearance is close to ordinary gaussian smoothing (honest — the fact that the type loosely passes yet the computation is essentially the same is shown as-is as a measured result).

1.7 op_sampler_3d.png — 3D op output sampler (bonus slot)

Generated by: tools/gen_article_assets.py::build_op_sampler_3d(). The data is the same as 1.3, studio_assets/sample_3d/itokawa_points.npy (the JAXA Hayabusa / Gaskell real point cloud). Six ops are run directly via ops3d.get(name).

op_sampler_3d

Panel op Meaning of the displayed numbers
Raw point cloud (input) The real 3000-point cloud shown as-is as a scatter (colored by distance from origin)
Point normals curvature3d.estimate_normals The normal vector of each point (principal-component estimation from k=20 neighbors) shown as arrows (thinned to 300 for legibility)
Shape index curvature3d.shape_index Colored by the Koenderink shape index (-1 = cup to +1 = cap). mean/std are a summary of the distribution
Voxel downsample pcl_filter.voxel_grid_downsample Grid thinning at a voxel size of (diameter/25). 3000 → 635 points (79% reduction)
Oriented bounding box pcseg.obb An oriented bounding box aligned with the principal inertia axes (red wireframe). extents = dimensions along the 3 axes (m)
Convex hull meshrepair.convex_hull The convex hull mesh of the point cloud (yellow wireframe). Vertex count and triangle count

2. examples_3d/_gallery/ — hero images / turntable GIFs of 3D examples

All figures are real data placed in examples_3d/_gallery/. The paths are relative to this page and are displayed as-is on GitHub.

Thumbnail File Description Generating script
render_beauty_hero.png Output of the hero renderer render_beauty, which composites all rendering quality layers (specular highlights, ambient occlusion, contact shadows, SSAA, tone mapping) in one shot (peanut-shaped mesh, metallic material) examples_3d/render_beauty.py
gear_hero.png Hero render of a spur gear (12 teeth) Cannot be identified (presumably a mesh derived from build_gear() in examples_3d/rotational_symmetry_fold.py, but the script that baked this image is not found in the current tree — noted here as honest disclosure)
hand_hero.png Hero render of a procedurally assembled whole-hand skeleton (8 carpals, 5 metacarpals, 14 phalanges, capsule SDF) Cannot be identified (presumably a mesh derived from build_hand_bones() in examples_3d/procedural_hand.py, but the script that baked this image is not found in the current tree)
fit_primitives_ext.png Extended primitive fitting that fits cones, tori, and triaxial ellipsoids to a point cloud (curvature-varying shapes that the existing RANSAC for cylinders/spheres/planes cannot handle) examples_3d/fit_primitives_ext.py
hull_bounds.png A set of primitives enclosing a point cloud (convex hull / AABB / OBB are re-uses of existing ops; the only new op is the minimum enclosing sphere min_enclosing_sphere) examples_3d/hull_bounds.py
mesh_decimate.png Boundary-preserving, strictly manifold mesh simplification by QEM edge-collapse (decimate_qem_manifold), compared empirically against naive random thinning and the existing op decimate_qem examples_3d/mesh_decimate.py
mesh_props.png Measurement of a triangle mesh’s normals, surface area, and mean curvature (discrete Laplace-Beltrami). Compared with the analytic solution of an icosphere (4πR²) examples_3d/mesh_props.py
mesh_smooth.png Smoothing of a noisy sphere mesh. Comparison of naive Laplacian smoothing (which shrinks) and Taubin smoothing (a non-shrinking band-pass filter) examples_3d/mesh_smooth.py
render_ao.png Ambient occlusion (selectively darkening contact areas and concavities by per-vertex hemispherical ray casting) examples_3d/render_ao.py
render_shade.png Phong specular highlights on a normal map and MatCap shading examples_3d/render_shade.py
render_shadow.png Cast shadows and penumbra via shadow mapping (soft shadows, area-light approximation) examples_3d/render_shadow.py
render_ssaa.png Removal of jaggies on mesh silhouettes via supersampling (SSAA) examples_3d/render_ssaa.py
render_tonemap.png Reinhard / ACES tone mapping of an HDR render (compared with a naive clip, demonstrating gradation preservation) examples_3d/render_tonemap.py
watershed3d.png Separating two touching objects by distance transform + watershed (a case where connected-component labeling fuses them into one) examples_3d/watershed3d.py
(GIF — animated on GitHub) showcase_turntable_pod.gif A turntable rotating the SDF-generated hero pod once with a metallic material tools/gen_showcase_gifs.py (build_pod)
(GIF) showcase_turntable_itokawa.gif The camera and sun orbit once around the real shape model of asteroid 25143 Itokawa (49,152 faces, no thinning) (phase angle 45°): adaptive tessellation 1.5 m, band-limited relief, angular rocks D^-3.1, Hapke, shadows from a solar angular diameter of 0.53°, zero ambient light tools/gen_itokawa_turntable.py
(GIF) showcase_turntable_skeleton.gif A hand-bone CT volume rotated once with a bone-color material (osteological-specimen style) tools/gen_showcase_gifs.py (build_skeleton, subject = hand-bone CT)
(GIF) showcase_hue_cycle.gif Rotating the hero pod while cycling the surface albedo hue from 0 to 360 tools/gen_showcase_gifs.py (build_pod + hue cycle)
(GIF) showcase_hand.gif Turntable of the procedural hand skeleton (presumed, the same subject as hand_hero.png) Cannot be identified (the regeneration script is not found in the current tree)

For the three items gear_hero.png / hand_hero.png / showcase_hand.gif, although the functions that generate the subjects (build_gear() / build_hand_bones()) themselves exist in the repository, the scripts that rendered them into images with render_beauty etc. and baked them into _gallery/ do not remain in the current tree. They are presumed to be ad-hoc runs from a past session (stated as a fact without asserting a guess).

3. Article thumbnails

In docs/articles/assets/thumbs/ there are thumbnails written out from the above montages, heroes, and new figures at 720px width (aspect preserved; not enlarged if the source is narrower than 720px). To keep file size down, they are saved as JPEG (quality=85, converted to RGB). The explanatory articles display these thumbnails, and the full sizes are referenced via this page.

4. Videos (mp4)

In docs/articles/assets/media/ there are H.264 mp4s written out from the same frames as the GIF showcases (light in size, playable as-is on GitHub’s blob pages), plus a video visualizing the event camera (DVS) stream of the Physical AI sensor family.

Video Description
pod.mp4 A turntable rotating the SDF-generated hero pod once with a metallic material (same frames as showcase_turntable_pod.gif)
itokawa.mp4 A physically based orbit around the real shape model of asteroid 25143 Itokawa (same frames as showcase_turntable_itokawa.gif, tools/gen_itokawa_turntable.py)
skeleton.mp4 A hand-bone CT volume rotated once with a bone-color material, osteological-specimen style (same frames as showcase_turntable_skeleton.gif)
hue_cycle.mp4 Rotating the hero pod while cycling the surface albedo hue from 0 to 360 (same frames as showcase_hue_cycle.gif)
dvs_stream.mp4 Event camera (DVS) simulation — while panning across a MuJoCo scene, the ON (bright) / OFF (dark) events generated flow along the object edges (the same log-luminance-difference model as event_camera.py run step by step; a lightweight dvs_stream.gif is also included)

Generated by: tools/gen_showcase_gifs.py::save_mp4() (four turntables, reusing the same frames as the GIFs) / tools/gen_article_assets.py::build_dvs_stream_video() (DVS stream).

5. Regenerate yourself

# Article montages + hero copies + 720px JPG thumbnails + DVS stream video (§1, §3, §4 of this page)
py -3.11 tools/gen_article_assets.py

# The 4 turntable GIFs in examples_3d/_gallery/ + the mp4s of the same frames (§2, §4 of this page,
# pod/itokawa/skeleton/hue_cycle). The script adds the repo root to sys.path itself, so
# setting PYTHONPATH is unnecessary.
py -3.11 tools/gen_showcase_gifs.py

# Individual 3D example hero images (examples)
py -3.11 examples_3d/render_beauty.py
py -3.11 examples_3d/render_ao.py
py -3.11 examples_3d/mesh_smooth.py
# ... for others, run examples_3d/<name>.py directly (each script overwrites _gallery/<name>.png)

# Studio screenshots (§6) / science gallery (§7) / cross-discipline (§8) / industrial + Physical AI (§9)
py -3.11 tools/gen_studio_screenshots.py
py -3.11 tools/gen_science_gallery.py
py -3.11 tools/gen_academic_gallery.py     # DL/generation cache is data/academic_samples/ (re-runs incur no billing, no re-DL)
py -3.11 tools/gen_industrial_gallery.py

6. Studio screenshots (studio_*.png)

Generated by: tools/gen_studio_screenshots.py. All are real screens headlessly grab()-ed from the actual Studio UI assembled by studio.build_window() (only the 3D surface uses Q3DSurface.renderToImage in a real GL context; the series construction shares Studio’s own _build_surface3d_series, built so that its appearance does not diverge from the real application). There are no mockups.

studio_main

Image Content
studio_main.png The main window. A blob segmentation pipeline (gaussian → otsu → opening_circle → sk_clear_border) is applied to the coins sample image, and 21 coins are overlaid in the region-overlay display. The lower Program panel shows HDevelop-style pipeline code, the right shows the operator browser (search + signature display), and the status bar shows 21 obj
studio_3d_surface.png The rotatable 3-D surface view opened with Ctrl+3 (Q3DSurface, height-linked terrain-style gradient). The data is the real-data relief of the Gaskell shape model of asteroid Itokawa (JAXA Hayabusa), depth-imaged via render3d.render_mesh. In the app, this view can be rotated by mouse drag and zoomed with the wheel
studio_python_editor.png The Python Editor (tabbed, editing multiple scripts simultaneously). Just after opening examples_3d/itokawa_curvature.py and running it with F5, the lower console shows the actual output of the Itokawa curvature analysis (PASS, exit 0)
studio_3d_examples.png The 3-D Examples gallery (105 real-data worked examples). Just after selecting itokawa_curvature and clicking Run, the Output tab shows the execution result with ground-truth verification (PASS)
studio_3d_ops.png The 3-D Operators reference (265 ops). The generated help page of icp_point2plane (signature, usage, verified samples, links to the next op where the types connect)

studio_3d_surface studio_python_editor studio_3d_examples studio_3d_ops

Generated by: tools/gen_science_gallery.py (regenerable per subject: py -3.11 tools/gen_science_gallery.py --subjects <name,...>). All are real outputs of fullseye’s registered ops / facade, with no mockups. Images derived from simulation state that fact explicitly in the caption. Thumbnails (720px-wide JPG) are the *_thumb.jpg in the same directory.

Image Content (ops used / data)
science_distance_ripple.png The distance transform of a real coin photo rendered with rainbow colors and ripple contours (otsu, fill_up, distance_transform / skimage.data coins)
science_fourier_stars.png The FFT spectrum of a real camera photo and a woven texture. The weave lights up like a constellation (fft_image; only the weave panel is synthetic)
science_watershed_foam.png 24 coins color-separated one by one with watershed (watersheds, segment_objects, colorize_labels)
science_edge_compass.png A neon image painting contour orientation with a hue wheel (sobel_dir, sobel_amp)
science_alife_worlds.png 6 panels of Rule 90 fractal / Rule 30 chaos / sandpile / DLA dendrite / Lenia / cyclic CA (iterated application of alife_*. Simulation images)
science_dino_xray.png A real scan of the Smithsonian Triceratops osteological specimen (CC0), voxelized → an X-ray look via vol_mip (voxelize, vol_gaussian, vol_mip)
science_dragon_anaglyph.png A red-cyan anaglyph of two-viewpoint renders of the Stanford dragon (read_mesh, look_at, render_mesh)
science_dino_terrain.png A 600k-point cloud of the osteological specimen → elevation map → terrain-shading coloring. The spine becomes a mountain range (elevation_map, colorize_height)
science_morph_pulse.gif A 26-frame morphology animation: merging by dilation → thinning by erosion (dilation_circle, erosion_circle)
science_wobble_warp.png Before/after of spatial deformation in the 3 styles of TPS/FFD/MLS (deform_tps/ffd/mls)
science_dino_skeleton.png Extracting the centerline in gold from the top-down silhouette of the osteological specimen (sk_skeleton)

science_dino_xray science_dragon_anaglyph

Generated by: tools/gen_academic_gallery.py. 30 exhibits = 8 real data + 22 AI-generated. It covers medicine, archaeology, biology, space, paleontology, geology, meteorology, oceanography, and botany, with every image structured as a before → after pair with fullseye’s registered ops applied.

Two representatives (for all 30, refer directly to articles/assets/academic_*.png):

academic_paleo_trex academic_arch_amphora

9. Industrial + Physical AI gallery (industrial_*.png / phai_*.png)

Generated by: tools/gen_industrial_gallery.py (--subjects selectable). All are real processing on synthetic data / MuJoCo simulations, and the detection/measurement results are confirmed with assert to match the known ground truth (placement count, drawn dimensions, placement pose).

Image Content (verification)
industrial_defect.png A hairline metal surface with 3 scratches, 2 dents, 1 foreign object → median background subtraction → red boxes + area (6/6 detected)
industrial_metrology.png Measuring the 3-step diameters of a stepped shaft with a sub-pixel caliper (max error 0.02px)
industrial_align.png Shape-matching localization of 3 rotated workpieces (position 0.0px, angle 0.0° match, no response to a different part)
industrial_blobs.png 60 pellets (6 touching pairs) counted with marker watershed 60/60 + size classification into 3 colors
industrial_barcode.png Edge-pair detection of 45 bars (all edges ±1.5px) + scan-line profile
phai_binpick.png Bin of parts from MuJoCo physical drop → height map → scoring 8 grasp candidates
phai_lidar_clusters.png 23k real ray casts → ground removal → 6 clusters + OBB bird’s-eye view (6/6)
phai_stereo_obstacles.png Disparity → 3D reconstruction → bird’s-eye obstacle map (4/4, median ground error 3mm)
phai_focus_stack.png 7 focal planes → all-in-focus composite (sharpness ×1.27)
media/phai_bin_pick.mp4 A full cycle where the Panda picks a grasp candidate, grabs it with 6-DOF IK, and carries it out (150 frames, 3/3 successful outfeed measured)

industrial_defect phai_binpick

This site is open source. Improve this page.