v0.3.0

What's Changed

Added

• Rust live-camera example (examples/yolov8_live.rs) — libcamera
  capture, zero-copy DMA-BUF tensor input, and direct Wayland DMA-BUF
  display for real-time YOLOv8 inference on NXP i.MX platforms.
• 3-step segmentation pipeline — split preprocessing, inference, and
  postprocessing into discrete HAL steps; reflected in both Rust and
  Python YOLOv8 examples.
• camera Cargo feature on the ara2 crate that gates the
  libcamera-based yolov8_live example behind optional dependencies
  (libcamera, wayland-client, wayland-protocols). Building the
  library and file-based examples no longer requires libcamera on the host.
• Monolithic YOLO detection decoder path for models with a single
  [1, nc+4, N] output tensor (previously only the split boxes+scores
  layout was supported).
• --format {nv12,yuyv} CLI flag on live-camera examples for pixel-format
  performance comparison (YUYV is ~1.3 ms faster than NV12 on imx95-frdm).
• --color-mode {class,instance,track} on all four YOLOv8 examples
  (previously hardcoded to Instance).
• --socket flag on examples/yolov8.rs for parity with the other examples.
• Comprehensive API documentation (doc comments / docstrings) across all
  example files.

Changed

• Live display backend: replaced EGL/GL with direct Wayland
  zwp_linux_dmabuf_v1 submission — no OpenGL context required, zero-copy
  from NPU output to compositor.
• Python camera capture: replaced GStreamer with native libcamera
  Python bindings in yolov8_live.py for lower latency and fewer
  transitive dependencies.
• Example CLIs: migrated Rust YOLOv8 examples to clap derive with
  per-variant --help descriptions and typo suggestions.

Fixed

• YOLO detection box mapping in examples/yolov8.rs: decoder produces
  normalized coordinates in the letterboxed model input frame (e.g. 640×640),
  not the original image. Un-pad and rescale by 1 / letterbox_scale so
  boxes are no longer stretched on non-square source images.

---

crates.io: ara2 | ara2-sys
PyPI: edgefirst-ara2