MENet

This is the official code for the paper 《MENet: Map-enhanced 3D object detection in bird’s-eye view for LiDAR point clouds》.

News

• 2024.02.14: The official code of MENet is published.
• 2023.05.19: 《MENet: Map-enhanced 3D object detection in bird’s-eye view for LiDAR point clouds》 is published on JAG.

Catalogue

• MENet
• News
• Catalogue
• Performance
• Installation
  • Tested Environment
  • Steps
• Dataset Preparation
• Train
  • Train on multiple GPUs
  • Train on single GPU
• Evaluate
  • Evaluate on multiple GPUs
  • Evaluate on single GPU
• Related Projects

Performance

Performance without CBGS: | Method | Modality | mAP | NDS | Weights | | :-----------------------------------------------: | :------: | :---: | :---: | :---------------------------------------------------------------------------------------------------: | | CenterPoint | L | 52.7 | 61.2 | Google Drive | | MENet | L+M | 56.9 | 63.4 | Google Drive | | SECOND | L | 35.7 | 51.3 | Google Drive | | ME SECOND | L+M | 43.1 | 55.5 | Google Drive |

Performance with CBGS: | Method | Modality | mAP | NDS | Weights | | :------------------------------------: | :------: | :---: | :---: | :-----------: | | CenterPoint | L | 56.2 | 64.7 | TODO | | MENet | L+M | 56.7 | 65.5 | TODO | | SECOND | L | 47.6 | 59.2 | TODO | | ME SECOND | L+M | 50.9 | 61.4 | TODO |

Installation

Tested Environment

• pyTorch
• mmcv-full==1.4.1
• mmsegmentation==0.14.1
• mmdet==2.14.0
• mmdet3d==0.17.1
• Pillow==8.4.0

Steps

1. Install the package listed above. We recommend that you create a new conda environment. To install the mmxxx series packages, you can click the hyperlinks and follow the instructions in the official documentations.
2. Install cuda extension.

python setup.py develop

Dataset Preparation

You can download, organize and prepare the dataset according to the documentory of mmdetection3d(nuScenes | Lyft).

Train

Train on multiple GPUs

export CUDA_VISIBLE_DEVICES=0,1,2,3 # 4 GPUs for example
export GPU_NUM=4
export CONFIG_FILE="configs/menet.py" # the config of any model
export WORK_DIR="./work_dirs/menet" # the output directory
./tools/dist_train.sh ${CONFIG_FILE}  ${GPU_NUM} --work-dir ${WORK_DIR} --deterministic

Train on single GPU

python tools/train.py ${CONFIG_FILE} --work-dir ${WORK_DIR} --deterministic

Evaluate

Evaluate on multiple GPUs

export CUDA_VISIBLE_DEVICES=0,1,2,3 # 4 GPUs for example
export GPU_NUM=4
export EVAL_METRICS=bbox

export CONFIG_FILE=./configs/menet.py
export CHECKPOINT_FILE="path of the weight of the model"
export RESULT_FILE=./work_dirs/menet/results.pkl

./tools/dist_test.sh ${CONFIG_FILE}  ${CHECKPOINT_FILE}  ${GPU_NUM} --out ${RESULT_FILE} --eval ${EVAL_METRICS}

Evaluate on single GPU

python tools/test.py ${CONFIG_FILE}  ${CHECKPOINT_FILE} --eval ${EVAL_METRICS}

Related Projects

• mmdetection
• BEVFusion