Release 0.1.0

What's New

• Initial public release of prism_image_proc: hardware-agnostic ROS 2 Humble image-processing accelerator with prism::ResizeNode as a drop-in replacement for image_proc::ResizeNode's resize pipeline.
• prism::ImageProcNode chainable base with action-chain composition (resize, crop, flip, colorconvert) and cv::resize direct-mode fallback when no GPU element validates.
• Thin component wrappers: prism::ResizeNode, prism::CropNode, prism::ColorConvertNode.
• Runtime hardware detection via HardwareDetector + live GStreamer registry validation via PipelineFactory (two-step probes per backend: nvvideoconvert → nvvidconv on Jetson, vapostproc → vaapipostproc on Intel).
• Zero-copy intra-process ingest (gst_buffer_new_wrapped_full + unique_ptr move), single-copy egress, no DDS round-trip on supported paths; verified empirically on Jetson Orin.
• A/B benchmark harness with Python NumPy stock baseline; A/B comparison launch file for side-by-side measurement against image_proc.
• Test/demo helpers: prism::MediaStreamerNode, prism::Synthetic4kPubNode.
• gtest unit suite covering the action-chain parser, all four CameraInfoTransform functors, and PipelineFactory.build().
• CI workflow on push and pull request (Humble container, gtest target).
• Project website at https://sohams25.github.io/prism-ros/ with architecture, parameters, benchmarks, and roadmap.
• Apache-2.0 license.

Known Limitations

• Intel VA-API: vaapipostproc chroma-subsampling regression on GStreamer 1.20 forces direct-mode fallback on stock Humble. GStreamer 1.22+ (Ubuntu 24.04 / Jazzy) is required for the GPU resize kernel on Intel.
• Jetson Orin (legacy nvvidconv): BGR-CAPS gap forces a CPU adapter on each pipeline boundary; per-action routing table sends crop to CPU videocrop. A JetPack image with the multimedia API installed (ships nvvideoconvert) closes this gap; capture is forward-work.
• Rectification not implemented (forward-work).