Release Notes: Surgical Instrument#

Version 2026.2.0#

Release Date: September 10, 2026

This is the initial release of the application. It is intended for reference and evaluation purposes only and not for direct use in clinical or diagnostic environments.

New:

  • Docker Compose driven workflow with a single make up entrypoint that supports both registry pull (default) and local source build via REGISTRY=false.

  • Decoupled capture / inference / display architecture: display shows every captured frame while OpenVINO inference runs on its own thread.

  • Pure OpenVINO detector (no ultralytics runtime coupling) with YOLO letterbox preprocessing and numpy NMS.

  • Multiple sources through a single interface: Basler camera (via pypylon), USB / V4L2 webcam, and video file.

  • OpenGL vsync presenter with cv2 and headless fallback.

  • Optional Basler capture trigger modes (off / software / vsync) with present-completion phase-lock for low photon-to-pixel latency.

  • Fullscreen OpenGL direct-scanout to bypass the compositor.

  • Per-stage latency CSV (trigger_to_grab, grab_to_display, trigger_to_display, infer_ms, disp_fps, cap_fps) for A/B comparison.

  • Optional core pinning and SCHED_FIFO real-time priority for capture, inference, and display threads.

  • Restructured layout with src/, docker/, and docs/ folders and a compose-only Makefile.

  • Optional local model-preparation flow: make setup-prerequisites installs the host prerequisites (Docker + Intel client GPU stack), then make backend-venv && make backend-bootstrap fetches the dataset, trains YOLO11n on the Intel iGPU, and exports a FP16 OpenVINO IR. See Model Preparation.

  • REAL-Colon dataset support: bootstrap auto-detects the REAL-Colon *_frames/ + *_annotations/ layout and converts Pascal VOC XML bounding boxes to YOLO labels. Legacy mask-based drops (ColonDB-style images+masks) remain supported as a fallback.

  • make check-l0 and make doctor preflight targets for host prerequisite verification.