v0.1.0 — Initial reference release

v0.1.0 — Initial reference release

First public release of the reference implementation.

What works

• SCRFD detector wrapper: three-stride decoder (8/16/32), NMS, letterboxed
  input, with output tensors resolved by shape so the stock insightface
  buffalo_l export runs directly.
• ArcFace ResNet50 embedder with batched TensorRT inference.
• 5-point face aligner using a Umeyama similarity transform.
• FAISS GPU index with adaptive IVF-Flat / IVF-PQ selection.
• Probe chain that batches detections across streams before encoding, and a
  multi-source DeepStream pipeline with thread-safe add_source /
  remove_source.
• Two CLI tools: face_enroll (build an index from a public dataset) and
  face_benchmark (per-stage latency / throughput), plus face_detect.
• 86 GoogleTest cases over the algorithmic stages — Umeyama transform, SCRFD
  decode, NMS, letterboxing, config validation, match decision — building
  without CUDA or TensorRT.

Known limitations

• The DeepStream probe parses tensor metadata for the SCRFD head only; a
  different detector needs its own decoder.
• Recognition is per-frame; there is no per-track confirmation layer.
• INT8 calibration for either engine is not validated end-to-end; FP16 is the
  documented configuration.
• The GPU stages (TensorRT wrapper, ArcFace encoder, FAISS searcher,
  DeepStream pipeline) are not unit tested — they need a device and a
  serialized engine.

Requirements

CMake 3.22+, a C++17 compiler, CUDA Toolkit 12.x, TensorRT 8.6+, DeepStream
SDK 7.x or 8.x, OpenCV 4.5+ with CUDA modules, Eigen 3.4+, spdlog, yaml-cpp,
FAISS (GPU build), gstreamer-1.0 headers. The Docker build under docker/
uses the official NVIDIA DeepStream devel image and is the most portable
route.

The unit tests need only CMake, a C++17 compiler, OpenCV, Eigen and yaml-cpp
(-DFACE_PIPELINE_CPU_ONLY=ON -DFACE_PIPELINE_BUILD_TESTS=ON).

This is a reference implementation intended to demonstrate architecture and
engineering patterns. See the README for scope and limitations.