Paper: https://arxiv.org/pdf/2207.02696.pdf
Special made for a Jetson Nano see Q-engineering deep learning examples
| Model | size | objects | mAP | Jetson Nano 1479 MHz | RPi 4 64-OS 1950 MHz |
|---|---|---|---|---|---|
| NanoDet | 320x320 | 80 | 20.6 | 26.2 FPS | 13.0 FPS |
| NanoDet Plus | 416x416 | 80 | 30.4 | 18.5 FPS | 5.0 FPS |
| YoloFastestV2 | 352x352 | 80 | 24.1 | 38.4 FPS | 18.8 FPS |
| YoloV2 | 416x416 | 20 | 19.2 | 10.1 FPS | 3.0 FPS |
| YoloV3 | 352x352 tiny | 20 | 16.6 | 17.7 FPS | 4.4 FPS |
| YoloV4 | 416x416 tiny | 80 | 21.7 | 16.1 FPS | 3.4 FPS |
| YoloV4 | 608x608 full | 80 | 45.3 | 1.3 FPS | 0.2 FPS |
| YoloV5 | 640x640 small | 80 | 22.5 | 5.0 FPS | 1.6 FPS |
| YoloV6 | 640x640 nano | 80 | 35.0 | 10.5 FPS | 2.7 FPS |
| YoloV7 | 412x412 tiny | 80 | 38.7 | 13.88 FPS | 4.43 FPS |
| YoloV7 | 640x640 tiny | 80 | 38.7 | 7.5 FPS | 2.1 FPS |
| YoloX | 416x416 nano | 80 | 25.8 | 22.6 FPS | 7.0 FPS |
| YoloX | 416x416 tiny | 80 | 32.8 | 11.35 FPS | 2.8 FPS |
| YoloX | 640x640 small | 80 | 40.5 | 3.65 FPS | 0.9 FPS |
To run the application, you have to:
- The Tencent ncnn framework installed. Install ncnn
- Code::Blocks installed. (
$ sudo apt-get install codeblocks)
To extract and run the network in Code::Blocks
$ mkdir MyDir
$ cd MyDir
$ wget https://github.com/Qengineering/YoloV7-ncnn-Jetson-Nano/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
YoloV7.cpb
yolo.cpp
yolo.h
yoloV7main.cpp
yolov7-tiny.bin
yolov7-tiny.param
To run the application load the project file YoloV7.cbp in Code::Blocks. More info or
if you want to connect a camera to the app, follow the instructions at Hands-On.
YoloV7 can handle different input resolutions without changing the deep learning model.
On line 28 of yolov7main.cpp you can change the target_size (default 640).
Decreasing the size to say 412 will speed up the inference time. On the other hand, the resizing makes the image less detailed; the model will no longer detect all objects.
Many thanks to nihui and Xiang Shin Wuu
