fullseye

Fullseye is a comprehensive, numpy-native operator library for computer vision, 3-D processing, and measurement. It provides over 1,500 typed operators (measured 2026-09-03: 870 distinct 2-D across 47 categories + 344 3-D across 63 categories) covering denoising, segmentation, morphology, edge detection, color conversion, point clouds, meshes, volumes, SDF, optics, tomography, interferometry, acoustics, and metrology. Every operator is reimplemented from published algorithms and open-source libraries (OpenCV, scikit-image, SciPy, Pillow, PyWavelets, SimpleITK, mahotas, kornia/torch), given a single typed interface, and contract-tested for finiteness, determinism, and sort-typing. On top of the operator library sits an evolutionary pipeline designer that evolves pipelines to maximize a user-defined metric on training data, with honest gating on a held-out set. The system also includes a sensor-simulation suite (pseudo-LiDAR, stereo, event camera, photometric stereo, TSDF fusion, polarization, focus stacking) for developing perception pipelines without hardware. Fullseye ships as an AI/RAG operator corpus: each operator has a machine-readable Markdown note (call form, type contract, HALCON counterpart, references) that serves as a retrieval corpus for AI coding assistants like Claude Code. A bundled Claude Code skill (`fullseye-rag`) sets up the RAG environment. An MCP server (PoC) provides 9 tools for LLM-driven operator search, reading notes, loading samples, running pipelines, and getting verdicts. A C ABI (`fullseye_abi.h`) with a generic entry point `fs_apply` reaches the whole 2-D registry (901 single-input operators) from C/C++/C#/Lua and Python ctypes, with a 5-operator contract that serves as a second implementation to find specification bugs. The library is pip-installable, requires only numpy and scipy, and degrades gracefully when optional backends (OpenCV, skimage, torch, etc.) are absent. GPU acceleration is opt-in via a batched torch fast path. Fullseye Studio is an HDevelop-style IDE with an operator browser, pipeline editor with breakpoints and timing, variable window, multi-window graphics, and a Python editor. The documentation includes a long-form overview (English and Japanese), a capabilities index, a maturity ledger, and design notes with 606 load-bearing comments in 6 languages. The project is archived on Zenodo from v0.1.11 onwards, with concept DOI 10.5281/zenodo.22761195. Licensed under Apache-2.0.