GETTING_STARTED.md

Getting Started

The dataset configs are located within cfgs/dataset_cfgs, and the model configs are located within cfgs for different settings.

Dataset Preparation

3D-FRONT Dataset

• Please download the subsampled and pre-processed 3D-FRONT dataset and organize it as follows.

  DODA
  ├── data
  │   ├── 3dfront
  │   │   │── density1250
  │   │   │── train_list.txt
  │   │   │── val_list.txt
  ├── cfgs
  ├── dataset
  

ScanNet Dataset

• Please download the ScanNet Dataset and follow PointGroup to pre-process the dataset as follows. Note that we change the ignore label index from -100 to 255.

  DODA
  ├── data
  │   ├── scannetv2
  │   │   │── train_group
  │   │   │   │── scene0000_00.pth
  │   │   │   │── ...
  │   │   │── val_group
  ├── cfgs
  ├── dataset
  

S3DIS Dataset

• Please download the S3DIS Dataset and follow PointNet to pre-process the dataset as follows or directly download the pre-processed data here.

  DODA
  ├── data
  │   ├── s3dis
  │   │   │── trainval_fullarea
  │   │   │   │── Area_1_Conference_1.npy
  │   │   │   │── ...
  ├── cfgs
  ├── dataset
  

Training and Inference

pretrain

sh scripts/train.sh ${NUM_GPUS} train --cfg_file ${CONFIG_FILE} ${PY_ARGS}

For instance, if you want to the train the pretrained model for 3D-FRONT $\rightarrow$ ScanNet with 8 GPUs:

sh scripts/train.sh 8 train --cfg_file cfgs/da_front3d_scannet/spconv.yaml

self-train

sh scripts/train.sh ${NUM_GPUS} st --cfg_file ${CONFIG_FILE} ${PY_ARGS}

For instance, if you want to the train the model for 3D-FRONT $\rightarrow$ ScanNet with 8 GPUs:

sh scripts/train.sh 8 st --cfg_file cfgs/da_front3d_scannet/spconv_st.yaml --weight output/da_front3d_scannet/spconv/default/ckpt/best_train.pth

Notice that you need to select the best model as your pretrain model, because the performance of adapted model is quite unstable.

Test a model

sh scripts/test.sh ${NUM_GPUS} --cfg_file ${CONFIG_FILE} ${PY_ARGS}

For instance, if you want to the test the model for 3D-FRONT $\rightarrow$ ScanNet with 8 GPUs:

sh scripts/test.sh 8 --cfg_file cfgs/da_front3d_scannet/spconv_st.yaml --ckpt output/da_front3d_scannet/spconv/default/default/ckpt/best_train.pth

Notice that you also need to focus on the performance of the best model.