According to lewes6369 's TensorRT-yolov3, a rewritten version of the TensorRT-yolov3 model that basically implements video and image reasoning and multi-threaded parallel acceleration has been successfully compiled on win10 system and Linux.
See my github for source code and compilation method .
Win10 version source code can be downloaded here .
build environment
ubuntu16 & win10
TensorRT 5.1
CUDA 9.0 or CUDA 10.0
Test effect
Model | GPU | Mode | Inference | FPS |
---|---|---|---|---|
yolov3_608 | GTX 1060 | fp32 | 58ms | 15 |
yolov3_608 | GTX 1060 | you8 | 38ms | 18 |
yolov3_608 | p40 | fp32 | 20ms | 40 |
yolov3_608 | p40 | you8 | 13ms | 50 |