Initial commit: scene-actor-extraction pipeline

Source (KPN++ pipeline nodes, ArcFace embedders, SCRFD/YuNet detectors,
gallery builder), build scripts, and eval artifacts.

- external/KPN as a git submodule (gitea.tourolle.paris/dtourolle/KPN)
- ONNX models tracked via Git LFS (models/*.onnx)
- generated outputs, TensorRT engines, reference repos, and media ignored

src/nodes/face_detector_node.hpp

@@ -0,0 +1,61 @@
+#pragma once
+#include "scrfd_decoder.hpp"
+#include "trt_scrfd_decoder.hpp"
+#include "config.hpp"
+#include "ort_provider.hpp"
+
+#include <memory>
+#include <string>
+
+// ── FaceDetectorFunc ──────────────────────────────────────────────────────────
+// KPN node: runs SCRFD-500MF to detect ALL faces in a frame.
+//
+// Backend selection:
+//   --detector-engine <path>  → TrtScrfdDecoder (raw TensorRT, no ORT)
+//   otherwise                 → SCRFDDecoder (ONNX Runtime)
+
+struct FaceDetectorFunc {
+    static constexpr std::string_view label() { return "face_detector"; }
+
+    explicit FaceDetectorFunc(const Config& cfg, OrtProvider provider)
+        : max_faces_(cfg.max_faces)
+        , min_face_px_(cfg.min_face_px)
+    {
+        if (!cfg.detector_engine.empty()) {
+            trt_ = std::make_unique<TrtScrfdDecoder>(
+                cfg.detector_engine, cfg.detector_conf, cfg.detector_nms);
+        } else {
+            ort_ = std::make_unique<SCRFDDecoder>(
+                cfg.detector_model, cfg.detector_conf, cfg.detector_nms, provider, cfg.trt);
+        }
+    }
+
+    SceneFrame operator()(Frame f) {
+        if (f.eof) return {std::move(f), {}};
+
+        auto faces = trt_ ? trt_->detect(f.image) : ort_->detect(f.image);
+
+        // Drop faces below minimum pixel size (too small for reliable ArcFace alignment)
+        faces.erase(
+            std::remove_if(faces.begin(), faces.end(), [&](const DetectedFace& d) {
+                return d.bbox.width < min_face_px_ || d.bbox.height < min_face_px_;
+            }),
+            faces.end());
+
+        // Sort largest-first so max_faces_ keeps the most informative detections
+        std::sort(faces.begin(), faces.end(),
+                  [](const DetectedFace& a, const DetectedFace& b) {
+                      return a.bbox.area() > b.bbox.area();
+                  });
+        if (static_cast<int>(faces.size()) > max_faces_)
+            faces.resize(max_faces_);
+
+        return {std::move(f), std::move(faces)};
+    }
+
+private:
+    std::unique_ptr<SCRFDDecoder>    ort_;
+    std::unique_ptr<TrtScrfdDecoder> trt_;
+    int                              max_faces_{10};
+    float                            min_face_px_{40.f};
+};