Attentive Prototypes for Source-free Unsupervised Domain Adaptive 3D Object Detection

PyTorch code release of the paper "Attentive Prototypes for Source-free Unsupervised Domain Adaptive 3D Object Detection",

by Deepti Hegde, Vishal M. Patel

[arXiv]

Follow the instructions for installation from ST3D

Dataset preperation

1. Download the relevant datasets: KITTI , Waymo , nuScenes

2. Generate the pickle files for each datset

• KITTI

  python -m pcdet.datasets.kitti.kitti_dataset create_kitti_infos tools/cfgs/dataset_configs/kitti_dataset.yaml
  

• nuScenes

  python -m pcdet.datasets.nuscenes.nuscenes_dataset --func create_nuscenes_infos --cfg_file tools/cfgs/dataset_configs/nuscenes_dataset.yaml  --version v1.0-trainval
  

• Waymo (this will take several hours for the whole dataset. You may download only a subset for a faster pre-processing and source training)

  python -m pcdet.datasets.waymo.waymo_dataset --func create_waymo_infos  --cfg_file tools/cfgs/dataset_configs/waymo_dataset.yaml
  

3. Organize each folder inside data like the following

   AttentivePrototypeSFUDA
   
   ├── data (main data folder)
   │   ├── kitti
   │   │   │── ImageSets
   │   │   │── training
   │   │   │   ├──calib & velodyne & label_2 & image_2 & (optional: planes)
   │   │   │── testing
   │   │   │   ├──calib & velodyne & image_2
   |
   |
   │   ├── nuscenes
   │   │   │── v1.0-trainval (or v1.0-mini if you use mini)
   │   │   │   │── samples
   │   │   │   │── sweeps
   │   │   │   │── maps
   │   │   │   │── v1.0-trainval  
   |
   |
   │   ├── waymo
   │   │   │── ImageSets
   │   │   │── raw_data
   │   │   │   │── segment-xxxxxxxx.tfrecord
   |   |   |   |── ...
   |   |   |── waymo_processed_data
   │   │   │   │── segment-xxxxxxxx/
   |   |   |   |── ...
   │   │   │── pcdet_gt_database_train_sampled_xx/
   │   │   │── pcdet_waymo_dbinfos_train_sampled_xx.pkl  
   
   

The below instructions are deprecated, please skip to this section for up-to-date instructions.

We implement the proposed method for two object detectors, SECOND-iou and PointRCNN for several domain shift scenarios. You can find the folder of pretrained models here. Find specific model downloads and their corresponding config files below.

| SECOND-iou |

| PointRCNN |

Follow the instructions to implement the method in the folders SECOND-iou and PointRCNN

Training

The entire training procedure may be divided into two stages: 1) source model training and 2) source-free domain adaptation.

Source model training

Single GPU training

  python tools/train.py --cfg_file tools/cfgs/da-waymo-kitti_models/secondiou/secondiou_cyc.yaml --extra_tag {PREFERRED NAME}

Multi-GPU training

  bash tools/scripts/dist_train.sh {NUM_GPUS} --cfg_file tools/cfgs/da-waymo-kitti_models/secondiou/secondiou_cyc.yaml --extra_tag {PREFERRED NAME}

Source-free domain adaptation

Choose the best performing model from the previous step and use it as the source trained model.

  python tools/train.py --cfg_file tools/cfgs/da-waymo-kitti_models/secondiou_attproto/secondiou_proto_ros_cyc.yaml --extra_tag {PREFERRED NAME} --pretrained_model {SOURCE_MODEL_PATH}

Testing

  python tools/test.py --cfg tols/cfgs/${PATH_TO_CONFIG_FILE} --extra_tag {PREFERRED_NAME} --ckpt {PATH_TO_CKPT}