PC-IGOS

Point Cloud Integrated-Gradients Optimized Saliency, a visualization technique that can automatically find the minimal saliency map that covers the most important features on a shape.

Authors

Ziwen Chen, Wenxuan Wu, Zhongang Qi

Publication

"Visualizing point cloud classifiers by curvature smoothing", in BMVC2020.

Usage

Preparing ModelNet40

1. Download and unzip the modelnet40_normal_resampled dataset to a data folder.
2. Create an empty checkpoint log folder for the classifier mkdir log.

Training the PointConv classifier

1. For training a PointConv point cloud classifier, run python3 train_pointconv.py.
2. If the training flattened, stop the process and tune down the lr hyperparam.

Precompute blurred shapes for ModelNet40 test split

1. Run python3 modelnetdataset.py. It will run the save_all_blurs() method defined there.

Evaluating PC-IGOS on ModelNet40 test split

1. Run python3 pc_IGOS.py. The main code for PC-IGOS is integrate_mask().
2. If you want to save point clouds along the del/ins curves, create a folder mkdir tensors and toggle visualize to be true in evaluate_on_all_classes().

Files

blur_utils.py : utils for smoothing curvatures on point clouds

modelnetdataset.py : PyTorch datasets for ModelNet40

pc_IGOS.py : main PC-IGOS code and evaluation code

pointconv.py : architecture of a PointConv classifier

pointconv_utils.py : utils for point cloud classifers

train_pointconv.py : training code for the PointConv classifier