feat: add YOLO auto-labeling with ONNX Runtime

[developer0hye]

Summary

Add auto-labeling capability using ONNX Runtime (CPU only) for Ultralytics YOLO models (.onnx). Users can load a pre-trained model and auto-detect objects on the current image or batch-process all images, then review and correct results using the existing manual annotation tools.

Key Changes

• New YoloDetector class (yolo_detector.h/cpp): Encapsulates ONNX Runtime inference with letterbox preprocessing, NMS postprocessing, and auto-detection of YOLOv5 ([B,N,C+5]) vs YOLOv8/11/12/26 ([B,C+4,N]) output formats. Reads ONNX metadata (class names, input size, task, end-to-end flag) for zero-config model loading.
• Auto-label toolbar in MainWindow: Load Model button, confidence threshold slider (1-99%), Auto Label (current image), Auto Label All (batch with progress dialog)
• Keyboard shortcut: R key for quick auto-labeling of the current image
• Undo support: Auto-label results can be undone with Ctrl+Z
• Class list mismatch warning: When the model's class list differs from the loaded label list, auto-label is disabled with a warning dialog explaining the mismatch
• Conditional compilation: #ifdef ONNXRUNTIME_AVAILABLE — app still builds and works without ONNX Runtime
• CI/CD: Downloads and bundles ONNX Runtime v1.24.1 pre-built binaries for Windows, Linux, and macOS
• CI accuracy test (Linux only): Validates C++ inference matches Ultralytics Python inference for YOLOv5n, YOLOv8n, YOLO11n, and YOLO26n using IoU-based detection matching
• Helper script: scripts/download_onnxruntime.sh for local development setup

Bug Fixes

• Fix out-of-bounds read in preprocess: constScanLine(y) already returns a pointer to row y, so the pixel offset should be x*3, not y*bytesPerLine + x*3. The old code doubled the y offset, reading past the image buffer for large images.
• Fix dynamic-shape ONNX models: Models with dynamic output dimensions (-1) are now handled correctly using runtime output shape instead of load-time shape.

Supported Models

Any Ultralytics model exported with model.export(format="onnx"):

• YOLOv5 (anchor-based, objectness score)
• YOLOv8, YOLO11, YOLO12, YOLO26 (anchor-free, same output format)
• End-to-end models (NMS baked in, [1, maxDet, 6] output)

CI Accuracy Test Results

All 4 model series pass with 100% detection match rate:

Model	Python	C++	Match	Result
YOLOv5n	2 dets	2 dets	100%	PASS
YOLOv8n	2 dets	2 dets	100%	PASS
YOLO11n	2 dets	2 dets	100%	PASS
YOLO26n	1 det	1 det	100%	PASS

How to Use

1. Open images and class list as usual
2. Click Load Model → select a .onnx file
3. Adjust confidence threshold with the slider
4. Click Auto Label (or press R) to detect objects on the current image
5. Click Auto Label All to batch-process all images
6. Review and correct results using existing annotation tools

🤖 Generated with Claude Code