TensorRT-CenterNet
demo (GT 1070)
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 |
- support Deform Conv v2.
- no nms.
- 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
Enviroments
- gtx 1070
pytorch 1.0-1.1
ubuntu 1604
TensorRT 5.0
onnx-tensorrt v5.0
cuda 9.0
- jetson TX2
jetpack 4.2
Models
- Convert CenterNet model to onnx. See here for details.
- Use netron to observe whether the output of the converted onnx model is (hm, reg, wh)
Example
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
