Follow instruction in OpenPCDet Install Guide
Carla3D (Simulation/Source) is available at not released yet.
TinySUSCape (already organized in KITTI format) is available from JST.
Please refer to 3D_adapt_auto_driving to prepare the KITTI and Lyft dataset. The train/val split txt of KITTI and Lyft used in our experiments is same as the 3D_adapt_auto_driving.
Build dataset info db with:
cd tools/
python -m pcdet.datasets.kitti.kitti_dataset create_kitti_infos cfgs/dataset_configs/{DATASET}.yaml- Train a source model on simulation domain (Carla3D):
cd tools/
python train.py --cfg_file cfgs/carla_models/pointrcnn_carla_org.yaml- Self-training with target domain dataset:
python train_st.py
--cfg_file cfgs/carla_models/pointrcnn_org_to_{DATASET}.yaml \
--mining_at 0 10 30 40 --mining_portion 0.3 0.5 0.7 1.0 \
# for lyft
# --mining_at 0 5 15 25 --mining_portion 0.3 0.5 0.7 1.0 \
--pretrained_model {BEST_MODEL_FROM_SOURCE_TRAIN} \python test.py \
--cfg_file cfgs/carla_models/pointrcnn_org_to_{DATASET}.yaml \
--batch_size 8 --workers 8 \
--ckpt {PATH_TO_CHECKPOINT}