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

TensorRT-Pelee can run over 200FPS on Titan V and 70FPS on jetson TX2(FP32)

Performance: 72 FPS, 13.8 ms (FP32) GPU Time:

Time by layers:

Requierments:

1.TensorRT4

2.CUDA 9.0

3.cudnn7

4.Jetpack 3.3(on TX2)

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