Awesome

[IJCAI 2024] Bridging Stereo Geometry and BEV Representation with Reliable Mutual Interaction for Semantic Scene Completion

Demo:

Benchmark Results

Table of Content

News
Quick Start
Installation
Prepare Data
Pretrained Model
Training & Evaluation
License
Acknowledgements

News

[2024/04]: Our new work is accepted on ECCV 2024, please check HTCL.
[2024/04]: Paper is accepted on IJCAI 2024
[2023/03]: Paper is on arxiv
[2023/03]: Demo and code released.

Quick Installation on A100

You can use our pre-picked environment on NVIDIA A100 with the following steps if using the same hardware:

a. Download the pre-picked package: occA100.

b. Unpack environment into directory occA100

cd /opt/conda/envs/
mkdir -p occA100
tar -xzf occA100.tar.gz -C occA100

c. Activate the environment. This adds occA100/bin to your path.

source occA100/bin/activate

You can also use Python executable file without activating or fixing the prefixes.

./occA100/bin/python

Step-by-step Installation Instructions

Following https://mmdetection3d.readthedocs.io/en/latest/getting_started.html#installation

a. Create a conda virtual environment and activate it. python > 3.7 may not be supported, because installing open3d-python with py>3.7 causes errors.

conda create -n occupancy python=3.7 -y
conda activate occupancy

b. Install PyTorch and torchvision following the official instructions.

conda install pytorch==1.10.1 torchvision==0.11.2 torchaudio==0.10.1 cudatoolkit=11.3 -c pytorch -c conda-forge

c. Install gcc>=5 in conda env (optional). I do not use this step.

conda install -c omgarcia gcc-6 # gcc-6.2

c. Install mmcv-full.

pip install mmcv-full==1.4.0

d. Install mmdet and mmseg.

pip install mmdet==2.14.0
pip install mmsegmentation==0.14.1

e. Install mmdet3d from source code.

cd mmdetection3d
git checkout v0.17.1 # Other versions may not be compatible.
python setup.py install

f. Install other dependencies.

pip install timm
pip install open3d-python
pip install PyMCubes

Known problems

AttributeError: module 'distutils' has no attribute 'version'

The error appears due to the version of "setuptools", try:

pip install setuptools==59.5.0

Prepare Data

a. You need to download
- The Odometry calibration (Download odometry data set (calibration files)) and the RGB images (Download odometry data set (color)) from KITTI Odometry website, extract them to the folder data/occupancy/semanticKITTI/RGB/.
- The Velodyne point clouds (Download data_odometry_velodyne) and the SemanticKITTI label data (Download data_odometry_labels) for sparse LIDAR supervision in training process, extract them to the folders data/lidar/velodyne/ and data/lidar/lidarseg/, separately.
b. Prepare KITTI voxel label (see sh file for more details)

bash process_kitti.sh

Pretrained Model

Download Pretrained model on SemanticKITTI and Efficientnet-b7 pretrained model, put them in the folder /pretrain.

Training & Evaluation

Single GPU

Train with single GPU:

export PYTHONPATH="."  
python tools/train.py   \
            projects/configs/occupancy/semantickitti/stereoscene.py

Evaluate with single GPUs:

export PYTHONPATH="."  
python tools/test.py    \
            projects/configs/occupancy/semantickitti/stereoscene.py \
            pretrain/pretrain_stereoscene.pth  1

Multiple GPUS

Train with n GPUs:

bash run.sh  \
        projects/configs/occupancy/semantickitti/stereoscene.py n

Evaluate with n GPUs:

 bash tools/dist_test.sh  \
            projects/configs/occupancy/semantickitti/stereoscene.py \
            pretrain/pretrain_stereoscene.pth  n

License

This repository is released under the Apache 2.0 license as found in the LICENSE file.

Acknowledgements

Many thanks to these excellent open source projects:

Citation

If you find our paper and code useful for your research, please consider citing:

@article{li2023bridging,
  title={Bridging stereo geometry and BEV representation with reliable mutual interaction for semantic scene completion},
  author={Li, Bohan and Sun, Yasheng and Liang, Zhujin and Du, Dalong and Zhang, Zhuanghui and Wang, Xiaofeng and Wang, Yunnan and Jin, Xin and Zeng, Wenjun},
  journal={arXiv preprint arXiv:2303.13959},
  year={2023}
}