TensorRT-YOLOv4 For TensorRT-7.1.0

Performance

Model	Input Size	GPU	Mode	Inference Time / FPS
YOLOv4	608x608	gtx 1080Ti	float32	23.3 ms
YOLOv4	416x416	gtx 1080Ti	float32	13.0 ms
YOLOv3	608x608	gtx 1080Ti	float32	18.2 ms
YOLOv3	416x416	gtx 1080Ti	float32	10.0 ms
YOLOv3-Tiny	608x608	gtx 1080Ti	float32	3.31 ms
YOLOv3-Tiny	416x416	gtx 1080Ti	float32	2.01 ms
YOLOv3-Tiny-Prn	608x608	gtx 1080Ti	float32	3.05 ms
YOLOv3-Tiny-Prn	416x416	gtx 1080Ti	float32	2.01 ms
YOLOv3	608x608	TX2 MAXN	float16	6 FPS
YOLOv4	416x416	TX2 MAXN	float16	7 FPS

• Note that including pre-processing and post-processing time.

Enviroments

• Jetson TX2

Ubuntu 18.04
TensorRT 7.1.0
CUDA 10.2
cuDNN 8.0.0
onnx 1.4.1
OpenCV 4.1.1
CMake 3.10.2

Models

• Modify [net] and [yolo] segments of yolovx.cfg.

[net]
batch=1
subdivisions=8
width=416
height=416
...

[yolo]
mask = 0,1,2
anchors = ...
down_stride = 8
classes=80
infer_thresh = 0.25
num=9
jitter=.3
ignore_thresh = .7
truth_thresh = 1
scale_x_y = 1.2
iou_thresh=0.213
...

[yolo]
mask = 3,4,5
anchors = ...
down_stride = 16
infer_thresh = 0.25
classes=80
num=9
jitter=.3
ignore_thresh = .7
truth_thresh = 1
scale_x_y = 1.1
iou_thresh=0.213
...

[yolo]
mask = 6,7,8
anchors = ...
classes=80
down_stride = 32
infer_thresh = 0.25
num=9
jitter=.3
ignore_thresh = .7
truth_thresh = 1
random=1
scale_x_y = 1.05
iou_thresh=0.213
...

• Convert darknet yolo to onnx.

python3 tools/yolo_to_onnx.py --cfg model/yolov4.cfg --weights model/yolov4.weights --out model/yolov4.onnx

Example

git clone https://github.com/hlld/tensorrt-yolov4.git
cd tensorrt-yolov4
mkdir build
cd build && cmake .. && make
cd ..

## YOLOv3
./buildEngine -i model/yolov3.onnx -o model/yolov3.engine -b 1 -m 0
./runDet -i model/yolov3.engine -p dog.jpg -v nuscenes_mini.mp4
### use web camera.
./videoDet -e model/yolov3.engine -c 1

## YOLOv4
./buildEngine -i model/yolov4.onnx -o model/yolov4.engine -b 1 -m 0
./runDet -i model/yolov4.engine -p dog.jpg -v nuscenes_mini.mp4
### use web camera.
./videoDet -e model/yolov4.engine -c 1

Related projects

• TensorRT-YOLOv4
• DarkNet
• TensorRT-CenterNet
• TensorRT
• onnx-tensorrt