Cloud-RAIN: Point Cloud Analysis With Reflectional-Invariance

This repository is an official implementation of the paper "Cloud-RAIN: Point Cloud Analysis With Reflectional-Invariance". Code will be released soon, stay tuned!

Tip: The result of point cloud experiment usually faces greater randomness than 2D image. We suggest you run your experiment more than one time and select the best result.

 

Point Cloud Semantic Segmentation on the S3DIS benchmark

You have to download Stanford3dDataset_v1.2_Aligned_Version.zip manually from https://goo.gl/forms/4SoGp4KtH1jfRqEj2 and place it under data/

Run the training script:

This task use 6-fold training, such that 6 models are trained leaving 1 of 6 areas as the testing area for each model.

• Train in area 1,2,3,4,6

python main_semseg_s3dis.py --exp_name=EXP_NAME --test_area=5 --model MODEL_NAME

Example:

python main_semseg_s3dis.py --exp_name=dgcnn_semseg_5_aug_no_norm --test_area=5 --model dgcnn

Run the evaluation script with pretrained models:

• Evaluate in area 5

python main_semseg_s3dis.py --exp_name=EXP_NAME--test_area=5 --model MODEL_NAME--eval True --model_root MODEL_ROOT

Example:

python main_semseg_s3dis.py --exp_name=dgcnn_semseg_5 --test_area=5 --model dgcnn --eval True --model_root ./checkpoints/dgcnn_semseg_5/models/

 

Point Cloud Semantic Segmentation on the ScanNet benchmark

Run the training script:

python main_semseg_scannet.py --exp_name=EXP_NAME --model MODEL_NAME

Example:

python main_semseg_scannet.py --exp_name=pointnet_semseg_scannet  --model pointnet