BEVMap - Map Aware BEV Modeling for 3D Perception

This is the official repository of the paper BEVMap: Map Aware BEV Modeling for 3D Perception.

Mincheol Chang, Seokha Moon, Reza Mahjourian, and Jinkyu Kim

Prerequisites

Installation

You can install environments as following commands. If error occurs, check torch and mmcv version compatibility, or you can refer to BEVDet installation guide.

conda create -n open-mmlab python=3.8 -y
conda activate open-mmlab
pip install torch==1.9.1+cu111 torchvision==0.10.1+cu111 torchaudio==0.9.1 -f https://download.pytorch.org/whl/torch_stable.html
pip install mmcv-full==1.4.0
pip install mmdet==2.14.0
pip install mmsegmentation==0.14.1
python setup.py develop 
pip install numba==0.48
pip install lyft_dataset_sdk
pip install nuscenes-devkit==1.0.5
pip install numpy==1.20.3
pip install yapf==0.40.1
pip install shapely==1.8.5

Data Preparation

You can directly follow BEVDet nuscenes_det.md to create pkl file for training and validation.

Or download nuscenes dataset and do following:

# symlink dataset root to ./data/
ln -s [nuscenes root] ./data/
# generate pickle files 
python tools/create_data.py nuscenes --root-path ./data/nuscenes --out-dir ./data/nuscenes --extra-tag nuscenes

Also download nuscenes map expansion and locate your data folders as following:

|-- data
|---|--- nuscenes
|---|---|--- v1.0-trainval/mini 
|---|---|--- samples 
|---|---|--- sweeps
|---|---|--- maps 
|---|---|---|--- basemap 
|---|---|---|--- boston-seaport.json
|---|---|---|--- ...

Map Dataset Generation

We provide codes for generating map projections and BEV map mask for every sample in nuScenes dataset.

Perspective View Map (RGB + Depth) Generation

python tools/mapdata/generate_projmap.py 
python tools/mapdata/generate_projdepth.py 

Example of augmented projected map and corresponding map depth

Projected Map	Projected Map with Map Depth

BEV Mask Generation

python tools/mapdata/generate_bevmask.py --coord 'lidar' --category 'drivable_area'

Note: We design the framework on master branch of BEVDet, which generates BEV features on LiDAR coordinates. However, recent dev2.0 branch creates BEV coordinates on egopose system, so map images should have corresponding coordinate transform too.

BEV Map on Ego Pose	BEV Map on LiDAR

Train

bash tools/dist_train.sh $CONFIG_FILE $GPUS --work_dirs $RESULT_PATH 

Inference

bash tools/dist_test.sh $CONFIG_FILE $CHECKPOINT $GPUS

Comparison with Baselines

FP32

• *trained without lidar supervision
• checkpoint will be released soon

Acknowledgement

Special thanks to Siyeong Lee from NAVER LABS for help with implementation of map generation codes and experimental feedbacks.

Our code is based on master branch of BEVDet. We thank and appreciate BEVDet teams for their great work.

We also thank other excellent open-source projects Lift-Splat-Shoot, mmdetection3d.