MMDetection3D-dockers

This is a docker environment for MMDetection3D repo. To get started, read their getting_started_doc

Prerequisites

Similar to carla official docker prerequisites

• Install docker following instructions here
• Install nvidia-docker2 following instructions here

MMDetection3D

Build the image:

docker build -t cisl/mmdet -f mmdet.Dockerfile .

To test, run

docker run -it --gpus all cisl/mmdet

Inside the docker, run

mim download mmdet3d --config pointpillars_hv_secfpn_8xb6-160e_kitti-3d-car --dest .

python demo/pcd_demo.py demo/data/kitti/000008.bin pointpillars_hv_secfpn_8xb6-160e_kitti-3d-car.py hv_pointpillars_secfpn_6x8_160e_kitti-3d-car_20220331_134606-d42d15ed.pth

mim download mmdet3d --config pointpillars_hv_secfpn_8xb6-160e_kitti-3d-car --dest . And check outputs/preds/000008.json

MMDetection3D with BEVFusion Extension and compilation

To build an MMdetection3D image with BEVFusion support:

docker build -t mmdet3d-bevfusion -f mmdet_bevfusion.Dockerfile .

To run the container:

docker run -it --gpus all --name mmdet3d_container mmdet3d-bevfusion bash

Verify BEVFusion by running the demo code on demo nuscenes data sample:

mkdir -p checkpoints
wget https://download.openmmlab.com/mmdetection3d/v1.1.0_models/bevfusion/bevfusion_lidar-cam_voxel0075_second_secfpn_8xb4-cyclic-20e_nus-3d-5239b1af.pth \
     -O checkpoints/bevfusion_lidar-cam_voxel0075_second_secfpn_8xb4-cyclic-20e_nus-3d-5239b1af.pth

python projects/BEVFusion/demo/multi_modality_demo.py \
    demo/data/nuscenes/n015-2018-07-24-11-22-45+0800__LIDAR_TOP__1532402927647951.pcd.bin \
    demo/data/nuscenes/ \
    demo/data/nuscenes/n015-2018-07-24-11-22-45+0800.pkl \
    projects/BEVFusion/configs/bevfusion_lidar-cam_voxel0075_second_secfpn_8xb4-cyclic-20e_nus-3d.py \
    checkpoints/bevfusion_lidar-cam_voxel0075_second_secfpn_8xb4-cyclic-20e_nus-3d-5239b1af.pth \
    --cam-type all