Mission: Deliver a production-grade, memory-safe, state-of-the-art fiducial marker detector.
Target Audience: Robotics, Autonomous Vehicles (AV), and Perception Engineers who require bounded latency and high reliability.
2. Engineering Directives
Latency Obsession: Every microsecond counts. Scrutinize cache lines, memory access patterns, and branching behavior. Prioritize Data-Oriented Design (DOD) over classical Object-Oriented patterns.
Concurrency First: All heavy pipeline stages must release the Python Global Interpreter Lock (GIL) to enable high-throughput multi-threaded perception systems.
Safety First: Rust's safety guarantees are a feature, not a hurdle. When bypassing them via unsafe (e.g., for SIMD or raw buffer access), the burden of proof is on the author to document soundness.
OpenCV Ecosystem Parity: We strictly adhere to modern OpenCV (cv2.aruco) conventions for tag layout, bit ordering (row-major), and canonical orientation. This ensures seamless interoperability with the broader computer vision ecosystem.
Visual Verifiability: Perception algorithms are notoriously difficult to debug via text. We mandate the use of the rerun SDK to emit rich, interactive visualizations of intermediate pipeline stages.
Modern Toolchain: We embrace the bleeding edge of tooling to improve developer velocity: uv for Python environments, cargo nextest for parallel testing, and maturin for seamless cross-language builds.
Performance Observability: We maintain nanosecond-level visibility into the pipeline without sacrificing production speed. By leveraging static tracing spans and compile-time erasure, we ensure that performance regressions are immediately detectable during profiling while maintaining zero runtime overhead in deployment.
Mathematical Rigor: Sub-pixel localization is a science of offsets. We strictly adhere to the +0.5 pixel center rule and center-aware decimation mapping to ensure that our synthetic benchmarks and real-world optimizations are numerically consistent with standard vision libraries (OpenCV/Blender).
Zero-Overhead Debugging: Our commitment to performance extends to our diagnostic tools. The Rerun-based debug pipeline follows a "pay-only-when-using" model: zero runtime impact on the production hot-path, arena-allocated ephemeral telemetry, and zero-copy data views.