gsplat op• Data kinds: table × matrix → table
• Call: import fullseye as fs; fs.ledger.gs_render(gs: 'dict', pose, K, width: 'int', height: 'int', *, light=(0.3, -0.5, 0.8), ambient: 'float' = 0.35, sky=(0.62, 0.75, 0.92), max_radius: 'int' = 64, max_pairs: 'int' = 12000000, antialias: 'bool' = True) -> 'dict' (to call the implementation directly, import gsplatnp; gsplatnp.gs_render(gs: 'dict', pose, K, width: 'int', height: 'int', *, light=(0.3, -0.5, 0.8), ambient: 'float' = 0.35, sky=(0.62, 0.75, 0.92), max_radius: 'int' = 64, max_pairs: 'int' = 12000000, antialias: 'bool' = True) -> 'dict'; from the registry, opsdrive.get("gs_render"))
Renders a 3DGS with a camera (pose = world→camera 4×4, K 3×3; same convention as driveworld.world_camera: looks along −Z, rows grow downward).
Steps: (1) centres into the camera frame, discard depth d = −z ≤ 1e-6, (2) 2-D covariance Σ' = J W Σ Wᵀ Jᵀ + 0.3 I, J = [[fx/d, 0, fx·x/d²], [0, −fy/d, −fy·y/d²]], (3) in a window of radius r = ⌈3 √λ_max(Σ')⌉ (≤ max_radius; those exceeding it are not discarded, r is clipped), per pixel α = min(0.99, o · exp(−½ Δᵀ Σ'⁻¹ Δ)) (α < 1/255 is discarded), (4) per pixel in depth order front to back C = Σ c_i α_i T_i, T_i = Π_{j<i}(1 − α_j), background = (1 − Σ α_i T_i) · sky. Colour is Lambert (same as world_camera: ambient + (1 − ambient)|n·l|, light in the camera frame). Everything is array operations (Gaussian × window-pixel pairs laid out flat, sorted by pixel → depth, T from per-group cumulative sums).
Returns: color (H,W,3), alpha (H,W) (= Σ α_i T_i), depth (H,W) (α-weighted depth / alpha; NaN where alpha < 0.5), label (H,W) (label of the Gaussian with the largest weight; −1 where alpha < 0.5), n_pairs, n_visible.
fail-closed: image size ≤ 0, max_radius < 1, or the number of pairs exceeding max_pairs (a too-slow render is not run silently) raise ValueError.
Empty input (gs is an empty dict) raises ValueError.
• Sample-data catalog (download URLs / licences) — 2-D uses skimage.data (BSD/public domain) plus synthetic images; 3-D lists download URLs for real data sources (Stanford, PDS, …).
• Operator provenance and references — the sources of the research/methods this op family came from.
• The canonical algorithm (author, year) and its uses are named in the family usage guide above.
• poc_kendama — py -3.11 examples/poc_kendama.py
table as input)course_layout · course_occupancy · course_contains · world_build · world_camera · world_move · lidar_scan · rss_longitudinal_same
gsplat)gs_from_world · gs_update · gs_render_fn · gs_read_file
*Provenance: gsplatnp.py — DRIVE operator registry. This per-op note is generated by tools/opdocs.py md (do not hand-edit).*
© 2026 Kazufumi Furuse — Fullseye operator documentation. Licensed under Apache-2.0.