Skip to content

eric612/Pelee-Seg-TensorRT

 
 

Repository files navigation

Pelee-TensorRT

Accelerate Pelee with TensorRT Pelee: A Real-Time Object Detection System on Mobile Devices (NeurIPS 2018)

TensorRT-Pelee can run over 70FPS(11ms) on Jetson TX2(FP32)


Performance:
Jetson TX2: 72 FPS, 13.2~11 ms (FP32)
Titan V: 200FPS, 5 ms (FP32)

Requierments:

1.TensorRT4 (Jetpack 3.3 on TX2)
2.CUDA 9.0
3.cudnn7


Run:

cmake .
make
./build/bin/pelee

Reference:

https://github.com/Ghustwb/MobileNet-SSD-TensorRT


TODO:

  • FP16 Implementation
  • Change Custom layers IPlugin to IPluginExt

The bug has been fixed

image

About

Pelee(NeurIPS'18)-TensorRT Implementation

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages

  • C++ 71.2%
  • Cuda 22.2%
  • C 3.4%
  • CMake 3.2%