Skip to content

Latest commit

 

History

History

detection3d

README.md

tools/detection3d

The pipeline to make the model. It contains training, evaluation, and visualization for 3D detection and 3D semantic segmentation.

  • Support priority: Tier S
  • Supported dataset
    • NuScenes
    • T4dataset with 3D detection
    • T4dataset with 3D semantic segmentation
  • Other supported feature
    • Add unit test

1. Setup environment

See setting environment

2. Prepare dataset

Prepare the dataset you use.

2.1. nuScenes

  • Run docker
docker run -it --rm --gpus '"device=0"' --shm-size=64g --name awml -p 6006:6006 -v $PWD/:/workspace -v $PWD/data:/workspace/data autoware-ml
  • Make info files for nuScenes
    • If you want to make own pkl, you should change from "nuscenes" to "custom_name"
python tools/detection3d/create_data.py nuscenes --root-path ./data/nuscenes --out-dir ./data/nuscenes --extra-tag nuscenes

2.2. T4dataset

  • Run docker
docker run -it --rm --gpus '"device=0"' --shm-size=64g --name awml -p 6006:6006 -v $PWD/:/workspace -v $PWD/data:/workspace/data autoware-ml
  • (Choice) Make info files for T4dataset XX1
    • This process takes time.
python tools/detection3d/create_data_t4dataset.py --root_path ./data/t4dataset --config autoware_ml/configs/detection3d/dataset/t4dataset/xx1.py --version xx1 --max_sweeps 2 --out_dir ./data/t4dataset/info/user_name
  • (Choice) Make info files for T4dataset X2
    • This process takes time.
python tools/detection3d/create_data_t4dataset.py --root_path ./data/t4dataset --config autoware_ml/configs/detection3d/dataset/t4dataset/x2.py --version x2 --max_sweeps 2 --out_dir ./data/t4dataset/info/user_name

3. Train

3.1. Change config

  • You can change batchsize by file name.
    • For example, 1×b1 -> 2×b8
  • If you use custom pkl file, you need to change pkl file from nuscenes_infos_train.pkl.

3.2. Train

  • Train in general by below command.
    • See each projects for detail command of training and evaluation.
python tools/detection3d/train.py {config_file}
  • You can use docker command for training as below.
docker run -it --rm --gpus '"device=0"' --name autoware-ml --shm-size=64g -d -v $PWD/:/workspace -v $PWD/data:/workspace/data autoware-ml bash -c '<command for each projects>'

3.3. Log analysis by Tensorboard

tensorboard --logdir work_dirs --bind_all

4. Analyze

4.1. Evaluation

  • Evaluation
python tools/detection3d/test.py {config_file} {checkpoint_file}
  • After training, you can test for each epoch checkpoints as below.
    • min_epoch is the epoch you want to start to test. If you set 20, test epoch_20.pth, epoch_21.pth, epoch_22.pth...
python tools/detection3d/test_all.py {config_file} {train_results_directory} {min_epoch}

4.2. Visualization

4.2.1 Rerun

  • Visualization for 3D view

See rerun_visualization

  • Visualization for BEV view
    • This script is simple debug tools.
    • This tool don't have image information, so you can use it when you make visualization images for public place like PR.
    • Image information often has personal information, so you cannot use it in public place in many case.
python tools/detection3d/visualize_bev.py {config_file} --checkpoint {config_file}

4.2.2 Custom visualization

  • Visualization for both BEV and image view
    • This tool is a simple visulization tool to visualize outputs from a 3D perception model
    • It needs to convert non-annotated T4dataset to an info file beforehand
    • For example, CenterPoint:
# Generate predictions for t4dataset
DIR="work_dirs/centerpoint/t4dataset/second_secfpn_2xb8_121m_base/" &&
python tools/detection3d/visualize_bboxes.py projects/CenterPoint/configs/t4dataset/second_secfpn_2xb8_121m_base.py $DIR/epoch_50.pth --data-root <new data root> --ann-file-path <info pickle file> --bboxes-score-threshold 0.35 --frame-range 700 1100

where ann-file-path is a path to the info file , and frame-range represents the range of frames to visualze

5. Deploy

See each projects