Use when working on XFeat pre/post-processing, keypoints, descriptors, or matching — tensor shapes, coordinate spaces, normalization, NMS/top-K, and the exact algorithm steps implemented in vrt-xfeat.
Use when working on multi-object tracking or metric-3D geometry — the vrt-track 3D Kalman, ByteTrack association, depth gating, NSA noise, occlusion recovery, coasting, world-frame BEV, camera intrinsics/extrinsics/undistort (vrt-types), or the…
Use when streaming the annotated pipeline to a browser/phone, or touching vrt-viz's H.264/WebSocket/WebCodecs live view (crates/vrt-viz: stream.rs, serve.rs, h264.rs, client.html) — encoder pipeline, WebCodecs client, jitter buffer, bitrate,…
Use when building or composing a real-time CV pipeline from vrt model crates — wiring RfDetrSeg / DepthAnything / XFeat together, running N models on one frame, feeding the GPU results to the CPU tracker (vrt-track) + live view (vrt-viz, H.264/WebCodecs),…
Use when adding a new model crate to vision-rt (a new detector/segmenter/depth/pose net) — crate layout, the async caller-owned pattern, classifying engine outputs by shape, from_engine_file/from_onnx/from_hub constructors, vrt-hub registration, export/build…
Use when writing or modifying CUDA code from Rust in this repo — kornia CudaKernel JIT, cudarc launches/memory/events, device pointers from TRT, or the trt-sys C FFI shim. Covers this repo's conventions, not general CUDA.
Use when building, rebuilding, or debugging TensorRT engines on this Jetson — shape profile errors ("does not satisfy any optimization profiles"), new ONNX exports, FP16 engines, or engine/version mismatch on deserialize.
Use when writing, optimizing, or reviewing CUDA kernel code itself (the __global__ functions) — block/grid sizing, memory coalescing, divergence, shared memory, precision, and Jetson Orin (SM87) specifics. Complements rust-cuda-patterns (which covers the Rust…