I test yolov4 on Jetson Nano B01(MAXIN mode - 10W) and JetPack 4.4(R32 4.3).
My work step:
- git clone GitHub - AlexeyAB/darknet: YOLOv4 / Scaled-YOLOv4 / YOLO - Neural Networks for Object Detection (Windows and Linux version of Darknet )
- yolov4.weights : https://drive.google.com/open?id=1cewMfusmPjYWbrnuJRuKhPMwRe_b9PaT
- modify Makefile : CUDA=1 , CUDNN=1, OPENCV=1
- modify cfg/yolov4.cfg : batch=1, subdivisions=1
- make
- ./darknet detect cfg/yolov4.cfg yolov4.weights data/test_image.jpg
I get result :
Done! Loaded 162 layers from weights-file
Detection layer: 139 - type = 27
Detection layer: 150 - type = 27
Detection layer: 161 - type = 27
data/dog.jpg: Predicted in 2870.691000 milli-seconds.
bicycle: 92%
dog: 98%
truck: 92%
pottedplant: 33%
" Predicted in 2870.691 ms" , Is this performance value normal?
Hi,
Since we are more familiar with pjreddie version, do you know the difference between this two GitHub source?
We try to run the YOLOv4 model with the yolov4.cfg in pjreddie version but get killed.
We are going to test the AlexeyAB source and update more information with you later.
Thanks.
1 Like
I’m able to run AlexeyAB’s darknet YOLOv4 416x416 model on Jetson Nano with JetPack-4.4. FPS is around 1.1. I have documented the result in my blog post: YOLOv4 on Jetson Nano
I have now implemented TensorRT YOLOv4. These are the performance numbers I got when I tested it on Jetson Nano with JetPack-4.4.
| TensorRT engine |
mAP @
IoU=0.5:0.95 |
mAP @
IoU=0.5 |
FPS on Nano |
| yolov4-288 (FP16) |
0.372 |
0.590 |
6.18 |
| yolov4-416 (FP16) |
0.454 |
0.698 |
3.50 |
| yolov4-608 (FP16) |
0.484 |
0.735 |
1.77 |
Update FPS numbers based on my latest implementation of tensorrt yolov3/yolov4:
