MonoGRNet: A Python repository from Yappedyen

MonoGRNet: A Geometric Reasoning Network for 3D Object Localization

Created by Zengyi Qin, Jinglu Wang and Yan Lu. The repository contains an implementation of this AAAI Oral Paper.

Video Demo | Detection Outputs on KITTI Validation Set

Please cite this paper if you find the repository helpful:

@article{qin2019monogrnet, 
  title={MonoGRNet: A Geometric Reasoning Network for 3D Object Localization}, 
  author={Zengyi Qin and Jinglu Wang and Yan Lu},
  journal={The Thirty-Third AAAI Conference on Artificial Intelligence (AAAI-19)},
  year={2019}
}

Prerequisites

Ubuntu 18.04
Python 3.6
Tensorflow 1.14

GPU-1080Ti

Install

Clone this repository

git clone https://github.com/Yappedyen/MonoGRNet.git

Download the Kitti Object Detection Dataset (image, calib and label) and place it into data/KittiBox. The folder should be in the following structure:

data
    KittiBox
        training
            calib
            image_2
            label_2
        train.txt
        val.txt

or you can follow this

cd MonoGRNet
sudo ln -s /path_to_kitti/training data/KittiBox/training

The train-val split train.txt and val.txt are contained in this repository.

Compile the Cython module and download the pretrained model:

python setup.py

Training and evaluation

Run the training script and specify the GPU to use:

python train.py --gpus 0

The evaluation is done during training. You can adjust the evaluation intervals in hypes/kittiBox.json. At the same time, you can open tensorboard see the graph and the processing of training. Open Terminal:

tensorboard --logdir=/home/.../MonoGRNet/outputs/kittiBox

After training, you can continue training after adjust the max_steps and opt in '/outputs/KittiBox/model_files/hypes.json'. Then use

python continue.py --logdir=/home/.../outputs/kittiBox --gpu=0

Visualization

cd visualize && mkdir visualize
python visualize.py

Acknowledgement

We would like to thank the authors of KittiBox for their code.

Yappedyen/MonoGRNet