Special made for the NPU, see Q-engineering deep learning examples
All models, with C++ examples can be found on the SD images.
All models are quantized to int8, unless otherwise noted.
| demo | model_name | RK3588 | RK3566/68 |
|---|---|---|---|
| yolov5 | yolov5s_relu | 50.0 | 14.8 |
| yolov5n | 58.8 | 19.5 | |
| yolov5s | 37.7 | 11.7 | |
| yolov5m | 16.2 | 5.7 | |
| yolov6 | yolov6n | 63.0 | 18.0 |
| yolov6s | 29.5 | 8.1 | |
| yolov6m | 15.4 | 4.5 | |
| yolov7 | yolov7-tiny | 53.4 | 16.1 |
| yolov7 | 9.4 | 3.4 | |
| yolov8 | yolov8n | 53.1 | 18.2 |
| yolov8s | 28.5 | 8.9 | |
| yolov8m | 12.1 | 4.4 | |
| yolov10 | yolov10n | 35.1 | 12.5 |
| yolov8s | 23.4 | 7.3 | |
| yolov8m | 9.7 | 3.4 | |
| yolov8x | 5.1 | 1.8 | |
| yolox | yolox_s | 30.0 | 10.0 |
| yolox_m | 12.9 | 4.8 | |
| ppyoloe | ppyoloe_s | 28.8 | 9.2 |
| ppyoloe_m | 13.1 | 5.04 | |
| yolov5_seg | yolov5n-seg | 9.4 | 1.04 |
| yolov5s-seg | 7.8 | 0.87 | |
| yolov5m-seg | 6.1 | 0.71 | |
| yolov8_seg | yolov8n-seg | 8.9 | 0.91 |
| yolov8s-seg | 7.3 | 0.87 | |
| yolov8m-seg | 4.5 | 0.7 | |
| ppseg | ppseg_lite_1024x512 | 27.5 | 2.4 |
| RetinaFace | RetinaFace_mobile3201 | 243.6 | 88.5 |
| RetinaFace_resnet50_3201 | 43.4 | 11.8 | |
| PPOCR-Det | ppocrv4_det2 | 31.5 | 15.1 |
| PPOCR-Rec | ppocrv4_rec3 | 35.7 | 17.3 |
1 Input size 320x320
2 Input size 480x480
3 Input size 48x320, FP16
- Due to the pixel-wise filling and drawing, segmentation models are relatively slow
To run the application, you have to:
- OpenCV 64-bit installed.
- Optional: Code::Blocks. (
$ sudo apt-get install codeblocks)
Start with the usual
$ sudo apt-get update
$ sudo apt-get upgrade
$ sudo apt-get install cmake wget curl
Follow the Raspberry Pi 4 guide.
$ git clone https://github.com/airockchip/rknn-toolkit2.git
We only use a few files.
rknn-toolkit2-master
│
└── rknpu2
│
└── runtime
│
└── Linux
│
└── librknn_api
├── aarch64
│ └── librknnrt.so
└── include
├── rknn_api.h
├── rknn_custom_op.h
└── rknn_matmul_api.h
$ cd ~/rknn-toolkit2-master/rknpu2/runtime/Linux/librknn_api/aarch64
$ sudo cp ./librknnrt.so /usr/local/lib
$ cd ~/rknn-toolkit2-master/rknpu2/runtime/Linux/librknn_api/include
$ sudo cp ./rknn_* /usr/local/include
Save 2 GB of disk space by removing the toolkit. We do not need it anymore.
$ cd ~
$ sudo rm -rf ./rknn-toolkit2-master
To extract and run the network in Code::Blocks
$ mkdir *MyDir* <br/>
$ cd *MyDir* <br/>
$ git clone https://github.com/Qengineering/YoloX-NPU.git <br/>
You can use Code::Blocks.
- Load the project file *.cbp in Code::Blocks.
- Select Release, not Debug.
- Compile and run with F9.
- You can alter command line arguments with Project -> Set programs arguments...
Or use Cmake.
$ cd *MyDir*
$ mkdir build
$ cd build
$ cmake ..
$ make -j4
Make sure you use the model fitting your system.
More info or if you want to connect a camera to the app, follow the instructions at Hands-On.
