Paper: https://arxiv.org/pdf/2405.14458
Special made for a bare Raspberry Pi 4, see Q-engineering deep learning examples
Numbers in FPS and reflect only the inference timing. Grabbing frames, post-processing and drawing are not taken into account.
| Model | size | mAP | Jetson Nano | RPi 4 1950 | RPi 5 2900 | Rock 5 | RK35881 NPU |
RK3566/682 NPU |
Nano TensorRT |
Orin TensorRT |
|---|---|---|---|---|---|---|---|---|---|---|
| NanoDet | 320x320 | 20.6 | 26.2 | 13.0 | 43.2 | 36.0 | ||||
| NanoDet Plus | 416x416 | 30.4 | 18.5 | 5.0 | 30.0 | 24.9 | ||||
| PP-PicoDet | 320x320 | 27.0 | 24.0 | 7.5 | 53.7 | 46.7 | ||||
| YoloFastestV2 | 352x352 | 24.1 | 38.4 | 18.8 | 78.5 | 65.4 | ||||
| YoloV2 20 | 416x416 | 19.2 | 10.1 | 3.0 | 24.0 | 20.0 | ||||
| YoloV3 20 | 352x352 tiny | 16.6 | 17.7 | 4.4 | 18.1 | 15.0 | ||||
| YoloV4 | 416x416 tiny | 21.7 | 16.1 | 3.4 | 17.5 | 22.4 | ||||
| YoloV4 | 608x608 full | 45.3 | 1.3 | 0.2 | 1.82 | 1.5 | ||||
| YoloV5 | 640x640 nano | 22.5 | 5.0 | 1.6 | 13.6 | 12.5 | 58.8 | 14.8 | 19.0 | 100 |
| YoloV5 | 640x640 small | 22.5 | 5.0 | 1.6 | 6.3 | 12.5 | 37.7 | 11.7 | 9.25 | 100 |
| YoloV6 | 640x640 nano | 35.0 | 10.5 | 2.7 | 15.8 | 20.8 | 63.0 | 18.0 | ||
| YoloV7 | 640x640 tiny | 38.7 | 8.5 | 2.1 | 14.4 | 17.9 | 53.4 | 16.1 | 15.0 | |
| YoloV8 | 640x640 nano | 37.3 | 14.5 | 3.1 | 20.0 | 16.3 | 53.1 | 18.2 | ||
| YoloV8 | 640x640 small | 44.9 | 4.5 | 1.47 | 11.0 | 9.2 | 28.5 | 8.9 | ||
| YoloV9 | 640x640 comp | 53.0 | 1.2 | 0.28 | 1.5 | 1.2 | ||||
| YoloV10 | 640x640 nano | 38.0 | 1.66 | |||||||
| YoloX | 416x416 nano | 25.8 | 22.6 | 7.0 | 38.6 | 28.5 | ||||
| YoloX | 416x416 tiny | 32.8 | 11.35 | 2.8 | 17.2 | 18.1 | ||||
| YoloX | 640x640 small | 40.5 | 3.65 | 0.9 | 4.5 | 7.5 | 30.0 | 10.0 |
1 The Rock 5 and Orange Pi5 have the RK3588 on board.
2 The Rock 3, Radxa Zero 3 and Orange Pi3B have the RK3566 on board.
20 Recognize 20 objects (VOC) instead of 80 (COCO)
YoloV10 use layer operations yet unknown to ncnn. To let YoloV10 run on an ncnn framework, zhouweigogogo has modified the onnx file by replacing these layers. At the same time, he expanded the original output to three. Some features, like variable input size, aren't supported now. Until the ncnn framework fully supports onnx opset 13, we face these limitations.
To run the application, you have to:
- A Raspberry Pi 4 or 5 with a 32 or 64-bit operating system. It can be the Raspberry 64-bit OS, or Ubuntu 18.04 / 20.04. Install 64-bit OS
- The Tencent ncnn framework installed. Install ncnn
- OpenCV 64-bit installed. Install OpenCV 4.5
- Code::Blocks installed (Optional). (
$ sudo apt-get install codeblocks)
To extract and run the network in Code::Blocks
$ mkdir MyDir
$ cd MyDir
$ wget https://github.com/Qengineering/YoloV10-ncnn-Raspberry-Pi-4/archive/refs/heads/main.zip
$ unzip -j master.zip
Remove master.zip, LICENSE and README.md as they are no longer needed.
$ rm master.zip
$ rm LICENSE
$ rm README.md
Your MyDir folder must now look like this:
parking.jpg
busstop.jpg
YoloV10.cpb
yoloV10.cpp
yolov10n.ncnn.bin
yolov10n.ncnn.param
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
if you want to connect a camera to the app, follow the instructions at Hands-On.
A more than special thanks to THU-MIG, who adapted the YoloV10 to onnx models.
And zhouweigogogo, who adapted the onnx model to the ncnn framework.
