feat(ANSODEngine): run Darknet/YOLO exports ("darknet" family) in ONNXYOLO and RTYOLO

A Darknet/YOLO model exported by hank-ai's darknet_onnx_export (v5.1+) has
two outputs: confs [B,N,nc], already objectness x class probability, and
boxes [B,N,1,4], corners x1,y1,x2,y2 normalized to the input. It trains and
predicts on a stretched square and has no NMS in the graph. ONNXYOLO took
the 4-D boxes tensor for segmentation proto masks and letterboxed the input,
so such a model decoded as garbage; a sealed package named as YOLOV4 went to
YOLOOD, which cannot open an .ansmodel at all.

- ONNXYOLO / RTYOLO: recognise the layout once at load (package family
  "darknet", output names confs/boxes, or their shapes), before the seg
  check; stretch-resize the input (single, batch, and RTYOLO's NV12 path);
  new postprocessDarknet: best class per candidate, threshold, stretch map
  to the original image, clip, class-aware NMS.
- dllmain: a sealed package passed as modelType 1 (YOLOV4) is routed as
  detection (3), i.e. to RTYOLO on NVIDIA and ONNXYOLO elsewhere, in
  CreateANSODHandle, CreateANSODHandleEx and OptimizeModelStr.
- Every other model is unchanged: all new branches are gated on the layout.

Verified on macOS only: ANSONNXYOLO.cpp passes a syntax check, and the new
decode functions, extracted into a standalone harness running the model
through onnxruntime, reproduce Darknet's own detections on dog.jpg. NOT yet
built on Windows, and the TensorRT path has not run.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
This commit is contained in:
2026-09-26 18:43:13 +10:00
co-authored by Claude Opus 5.5
parent cb964907f0
commit 0baca1b20d
6 changed files with 378 additions and 8 deletions
+22
View File
@@ -19,6 +19,7 @@ namespace ANSCENTER {
// - Segmentation: [B,300,38]+protos or [B, nc+36, N]+protos
// - Pose : [B,300,6+nk*3] or [B, nc+4+nk*3, N] legacy
// - Classification: [B, nc]
// - Darknet detect: confs [B,N,nc] + boxes [B,N,1,4] (stretch resize)
//
// Preprocessing follows the exact Ultralytics LetterBox transform:
// - center=True, stride=32, pad_value=114
@@ -95,6 +96,14 @@ namespace ANSCENTER {
/// True if the ONNX model has dynamic spatial dimensions.
bool hasDynamicInputShape() const { return isDynamicInputShape; }
/// Darknet/YOLO layout (hank-ai darknet_onnx_export): outputs
/// confs [B,N,nc] + boxes [B,N,1,4], stretch resize, no letterbox.
/// Detected from the output names/shapes at construction; the
/// wrapper also forces it when an .ansmodel package says its
/// family is "darknet".
void setDarknetLayout(bool darknet) { isDarknetLayout = darknet; }
bool hasDarknetLayout() const { return isDarknetLayout; }
private:
Ort::Value transform(const cv::Mat& mat) override;
Ort::Value transformBatch(const std::vector<cv::Mat>& images) override;
@@ -155,6 +164,13 @@ namespace ANSCENTER {
const std::vector<std::string>& classNames,
float confThreshold, float iouThreshold, int numKPS, int maxDet = 300);
// ── Darknet postprocess (confs + boxes, no NMS in the graph) ────
std::vector<Object> postprocessDarknet(
const cv::Size& originalImageSize,
std::vector<Ort::Value>& outputTensors,
const std::vector<std::string>& classNames,
float confThreshold, float iouThreshold, int maxDet = 300);
// ── Classification postprocess ──────────────────────────────────
std::vector<Object> postprocessClassify(
std::vector<Ort::Value>& outputTensors,
@@ -189,6 +205,12 @@ namespace ANSCENTER {
cv::Size inputImageShape;
bool isDynamicInputShape{ false };
// Darknet layout: decided once at load time (see setDarknetLayout).
// Selects the stretch resize in transformInto / transformBatchInto
// and postprocessDarknet in detect / detectBatch.
bool isDarknetLayout{ false };
void detectDarknetLayout();
// Per-call preprocess that writes into caller-provided buffer + dims
// instead of the BasicOrtHandler members (input_values_handler /
// input_node_dims). detect() / detectBatch() use these so multiple