Video

This repository is the official implementation of our paper SMOKE: Single-Stage Monocular 3D Object Detection via Keypoint Estimation. For more details, please see our paper.

SMOKE is a real-time monocular 3D object detector for autonomous driving. The runtime on a single NVIDIA TITAN XP GPU is ~30ms. Part of the code comes from CenterNet, maskrcnn-benchmark, and Detectron2.

The performance on KITTI 3D detection (3D/BEV) is as follows:

The pretrained weights can be downloaded here.

All codes are tested under the following environment:

• Ubuntu 16.04
• Python 3.7
• Pytorch 1.3.1
• CUDA 10.0

We train and test our model on official KITTI 3D Object Dataset. Please first download the dataset and organize it as following structure:

kitti
│──training
│    ├──calib 
│    ├──label_2 
│    ├──image_2
│    └──ImageSets
└──testing
     ├──calib 
     ├──image_2
     └──ImageSets

1. We use conda to manage the environment:

conda create -n SMOKE python=3.7

2. Clone this repo:

git clone https://github.com/lzccccc/SMOKE

3. Build codes:

python setup.py build develop

4. Link to dataset directory:

mkdir datasets
ln -s /path_to_kitti_dataset datasets/kitti

First check the config file under configs/.

We train the model on 4 GPUs with 32 batch size:

python tools/plain_train_net.py --num-gpus 4 --config-file "configs/smoke_gn_vector.yaml"

For single GPU training, simply run:

python tools/plain_train_net.py --config-file "configs/smoke_gn_vector.yaml"

We currently only support single GPU testing:

python tools/plain_train_net.py --eval-only --config-file "configs/smoke_gn_vector.yaml"

CenterNet

maskrcnn-benchmark

Detectron2

Please cite our paper if you find SMOKE is helpful for your research.

@article{liu2020SMOKE,
  title={{SMOKE}: Single-Stage Monocular 3D Object Detection via Keypoint Estimation},
  author={Zechen Liu and Zizhang Wu and Roland T\'oth},
  journal={arXiv preprint arXiv:2002.10111},
  year={2020}
}