Installation

pip install -e ops/point2mask 

Training Procedure

1. Follow the instructions under selected dataset (currently only ScanNetV2 is supported) to prepare data.
2. Run CUDA_VISIBLE_DEVICES=$SELECTED_DEVICE$ python -u train.py --config config/$CONFIG_NAME$ > $LOGFILENAME$.log 2>&1 & to train the network.
3. After training, you can first run python -u statistics.py to find out the appropriate logits threshold, and then run python -u pseudoLabelGeneration.py --config config/$CONFIG_NAME$to generate pseudo labels.
4. After that, prepare a new config specifying the path of pseudo label. See the existing config for reference.

You can train a bigger/more accurate network by changing m / block_reps / residual_blocks / scale / val_reps in config, e.g.

m: 32 # Wider network
block_reps: 2 # Deeper network
residual_blocks: True # ResNet style basic blocks
scale: 50 # 1/50 m :  2cm voxels
val_reps: 3 # Multiple views at test time
batch_size: 5 # Fit in 16GB of GPU memory

Switch Backbone

The followings are backbones available now.

• Point cloud encoder
  1. SparseConvUNet
  2. SparseConvFCNet
  3. SparseConvFCNetEncoder
• Text encoder
  1. TextTransformer
• Architecture
  1. MultiLabelContrastive
  2. MultiLabel

If you want to add new module, please register your module under models/ and specity embed_length if it is a point cloud encoder. Then switch the backbone's name in the config.

Note that the input should be [coords, feats] and the output should be logits for every points. If you are using a SparseConvNet-based point cloud encoder, just inherit from the class SparseConvBase_ and redefine the structure under getEncoder(*args). The rest will be handled automatically.