kaixinbear / rtm

my implementation of RTM3D ,though the precision is low

my implementation of RTM3d based on CenterNet

I have to emphasize that the precision is low and you'd better refer to official code for good performance.

Environment

Python3.6+ torch 0.4.1 following the centernet. Link-->env prepare

Installation

First,you need to prepare data like the centernet do. Link-->kitti data prepare

In order to compile src/lib/utils/energy.cpp,you need to install pybind11. Link-->pybind11

cd ~/Project/RTM/src/lib/utils

make 

 ~/PScanning dependencies of target energy
[ 50%] Building CXX object CMakeFiles/energy.dir/energy.cpp.o
[100%] Linking CXX shared module energy.cpython-36m-x86_64-linux-gnu.so
[100%] Built target energyroject/RTM/src/lib/utils$ make

Then test if you have install energy module successfully.

(CenterNet) kaixin1@213c6db174e2:~/Project/RTM/src/lib/utils$ python
Python 3.6.10 |Anaconda, Inc.| (default, Jan  7 2020, 21:14:29) 
[GCC 7.3.0] on linux
Type "help", "copyright", "credits" or "license" for more information.
>>> import energy
>>> 

##Train

python main.py ddd --exp_id 3dop --dataset kitti --kitti_split 3dop --batch_size 4  --num_epochs 70 --lr_step 45,60 --arch resFP_18

##Test

CUDA_VISIBLE_DEVICES=0 python test.py ddd --exp_id 3dop --dataset kitti --kitti_split 3dop --load_model ../models/model_180.pth --arch resFP_18 --gpus 3