| page_type | languages | products | name | urlFragment | |||
|---|---|---|---|---|---|---|---|
sample |
|
|
DirectML ONNX YOLOv9 NPU sample |
DirectML-YOLOv9 |
Implements the YOLOv9 real-time object detection model using DirectML, DirectMLX and Onnx runtime with the Qualcomm Hexagon NPU (and other NPU's?)
YOLOv9 is an object detection model capable of recognizing up to 80 different classes of objects in an image. This sample contains a complete end-to-end implementation of the model using DirectML, and is able to run in real time on a user-provided video stream.
This sample requires an external download of the ONNX YOLOv9 model, which are roughly 50MB large. The .vcxproj file is configured to download the file automatically, but it can also be downloaded manually here:
- https://drive.usercontent.google.com/download?id=1Yzu-vlH0mw2Jz8yBkkh12xuoDEDesDhN
- The model Model_Yolo_v9c_f16.onnx is quantized for the Hexagon NPU by:
- https://github.com/DakeQQ/YOLO-Depth-Estimation-for-Android
The file should be placed in a "Data" subdirectory. e.g.
<repo root>/Data/Model_Yolo_v9c_f16.onnx
Some more models are located in the Data folder in the release files.
Now supporting yolo segment models:
This sample demonstrates real time perfromance on a Snapdragon X Elite Dev Box. Compared to the original YOLOv4 DirectML Sample running on the GPU, this implementation runs more than twice as fast using less than half the power (15W compated to 40W on the GPU)
- DirectML NPU support (Drivers needed): https://blogs.windows.com/windowsdeveloper/2024/08/29/directml-expands-npu-support-to-copilot-pcs-and-webnn/
- Paper: YOLOv4: Optimal Speed and Accuracy of Object Detection (arXiv)
- Original GitHub: https://github.com/AlexeyAB/darknet
- TensorFlow implementation: https://github.com/hunglc007/tensorflow-yolov4-tflite