townim-faisal / lwf-3D

[IWANN 2021] Reducing catastrophic forgetting in 3D point cloud objects with help of semantic information

Learning without Forgetting for 3D Point Cloud Objects

Requirements

Install necessary packages from requirements.txt file.

Data

Semantic embedding for the dataset will be found here. You will find the class split in the paper.

Model

Pretrained model of old task: here

Pretrained model of new task: here

Training and Evaluation

For each dataset, there is a corresponding configuration files located in config folder. Below is the description of configuration file.

seen_class : number of classes for old task
unseen_class : number of classes for new task
total_class : number of total classes
dataset_path : path of the dataset i.e. "content/ModelnetNew"
saved_model : folder to save model for new task
batch_size : batch size
lr : learning rate
wd : weighting decay
T: temperature for KD loss
pointnet_old_model_path_none: model path for old task using pointnet (no semantic information)
pointnet_old_model_path_w2v: model path for old task using pointnet and word2vec
pointnet_old_model_path_glove: model path for old task using pointnet and glove
pointconv_old_model_path_none: model path for old task using pointconv (no semantic information)
pointconv_old_model_path_w2v: model path for old task using pointconv and word2vec
pointconv_old_model_path_glove: model path for old task using pointconv and glove
dgcnn_old_model_path_none: model path for old task using dgcnn (no semantic information)
dgcnn_old_model_path_w2v:  model path for old task using dgcnn and word2vec
dgcnn_old_model_path_glove: model path for old task using dgcnn and glove

For training and evaluating, arguments for each python script are:

--dataset: ModelNet, ScanObjectNN
--epoch: number of epochs 
--sem: using semantic representation i.e. w2v, glove, none

Acknowledgements

This implementation has been based on these repositories: PointNet, PointConv and DGCNN.

Citation

@inproceedings{lwf3D2021,
  title={Learning without Forgetting for 3D Point Cloud Objects},
  author={Townim Chowdhury, Mahira Jalisha, Ali Cheraghian, and Shafin Rahman},
  booktitle = {International Work-Conference on Artificial Neural Networks (IWANN)},
  year={2021}
}