TensorRT-CenterNet

This repo was originally forked from https://github.com/CaoWGG/TensorRT-CenterNet. Best to use nvidia Docker from NGC. Here is a summary:

1. nvidia-docker: cuda:10.0-cudnn7-devel-ubuntu18.04
2. Make sure to install: apt-get install -y wget vim tree libsm6 libxext6 libxrender-dev software-properties-common libgtest-dev libprotobuf-dev protobuf-compiler libopencv-dev swig gcc
3. Install TRT 7.0.0
4. GPU TITAN X

I will be replicating this work on Jetson and Jetpack 4.4

• ctdet_coco_dla_2x
  • 
• centerface
  • 
• cthelmet
  • 

Performance

model	input_size	GPU	mode	inference Time
mobilenetv2	512x512	gtx 1070	float32	3.798ms
mobilenetv2	512x512	gtx 1070	int8	1.75ms
mobilenetv2	512x512	jetson TX2	float16	22ms
dla34	512x512	gtx 1070	float32	24ms
dla34	512x512	gtx 1070	int8	19.6ms
dla34	512x512	jetson TX2	fp32	209ms
dla34	512x512	jetson TX2	fp16	186ms
dla34v0	512x512	gtx 1070	float32	12.6ms
dla34v0	512x512	gtx 1070	int8	6.76ms
dla34v0	512x512	jetson TX2	fp32	114ms
dla34v0	512x512	jetson TX2	fp16	80ms
resdcn101	512x512	gtx 1070	float32	20.9ms
resdcn18	512x512	gtx 1070	float32	5.81ms
resdcn18	512x512	gtx 1070	int8	3.63ms
resdcn18	512x512	jetson TX2	fp32	54ms
resdcn18	512x512	jetson TX2	fp16	41ms

1. support Deform Conv v2.
2. no nms.
3. support fp32 fp16 int8 mode.

Eval Result

model	GPU	mode	APtrt/APpaper	APtrt50	APtrt75	APtrtS	APtrtM	APtrtL
ctdet_coco_dla_2x	gtx 1070	float32	0.365/0.374	0.543	0.390	0.164	0.398	0.536
ctdet_coco_dlav0_1x	gtx 1070	float32	0.324/--	0.511	0.343	0.140	0.350	0.476
ctdet_coco_dlav0_1x	gtx 1070	int8	0.295/--	0.468	0.311	0.123	0.318	0.446
ctdet_coco_resdcn101	gtx 1070	float32	0.332/0.346	0.516	0.349	0.115	0.367	0.531
ctdet_coco_resdcn18	gtx 1070	float32	0.277/0.281	0.448	0.286	0.083	0.290	0.454
ctdet_coco_resdcn18	gtx 1070	int8	0.242/0.281	0.401	0.250	0.061	0.255	0.409

notes

• cocoval2017 test AP with no augmentation.
• input_szie = 512x512
• thresh = 0.01
• maxpool kernel_size = 3
• calib_img_list.txt : random sample 700 images from COCO2017/val2017

### Models
1. Convert [CenterNet](https://github.com/xingyizhou/centernet) model to onnx. See [here](readme/ctdet2onnx.md) for details.
2. Use [netron](https://github.com/lutzroeder/netron) to observe whether the output of the converted onnx model is (hm, reg, wh)

### Example
```bash
git clone https://github.com/CaoWGG/TensorRT-CenterNet.git
cd TensorRT-CenterNet
mkdir build
cd build && cmake .. && make
cd ..

##ctdet | config include/ctdetConfig.h 
## float32
./buildEngine -i model/ctdet_coco_dla_2x.onnx -o model/ctdet_coco_dla_2x.engine 
./runDet -e model/ctdet_coco_dla_2x.engine -i test.jpg -c test.h264

##cthelmet   | config include/ctdetConfig.h
## flaot32
./buildEngine -i model/ctdet_helmet.onnx -o model/ctdet_helmet.engine -m 0
./runDet -e model/ctdet_helmet.engine -i test.jpg -c test.h264
## int8
./buildEngine -i model/ctdet_helmet.onnx -o model/ctdet_helmet.engine -m 2 -c calib_img_list.txt
./runDet -e model/ctdet_helmet.engine -i test.jpg -c test.h264

##centerface | config include/ctdetConfig.h 
./buildEngine -i model/centerface.onnx -o model/centerface.engine
./runDet -e model/centerface.engine -i test.jpg -c test.h264

## run eval_coco.py | conifg your cocodaset and ctdet_coco engine 
python3 eval_coco.py model/ctdet_coco_dla_2x.engine

Related projects

• TensorRT-Yolov3
• onnx-tensorrt
• TensorRT
• CenterNet
• centerface
• netron
• cpp-optparse