Robust Object Classification Approachusing Spherical Harmonics

code for the paper: Robust Object Classification Approachusing Spherical Harmonics

This repository contains tensorflow implementation of a robust spherical harmonics CNN. The Method was tested for both robust point cloud and image (MNIST) classification. The model uses the voxel grid of concentric spheres to learn features over the unit ball. Also the convolution operations is performed in the Fourier domain. As a result, our model is able to learn features that are less sensitive to Point clouds perturbations and corruptions.

Training

Our code is based on spherical-cnn. The training and testing files should be stored as .tfrecord. We provide an examble on how to generate a .tfrecord from .h5. We perform the data augmentations in Matlab.

Point cloud

To generate the .tfrecord files:

1. Download ModelNet40 dataset, or any point cloud dataset with .h5 extension.
2. Use the makeSphVoxels.m to convert point cloud 3D Models to spherical voxel grid.
3. Use tfrecord_generator.py to generate the .tfrecord files.

After generating the training and testing files, run the Point cloud model as:

python3 scripts/train.py \
                               @params/model-64.txt \
                               @params/m40-64.txt \
                               @params/training.txt \
                               --dset_dir ~/data/m40-so3-64 \
                               --logdir /tmp/m40-so3 \
                               --run_id m40-so3

mnist

We followed the same procedure for generating the Corrupted mnist dataset. Refer to makemnist.m for details on how to generate the data corruptions. We also provide the data used in our experiments and a checkpoint inside the tmp folder.

We compared our method with a convertional CNN and two other spherical CNNs in the table below. Each mnist input image has 28*28 pixels, we randomly currupt 100 or 300 pixels as shown in the image above. The results are shown in the table below:

Method	clean	100	300
2D CNN	98	83.3	37.6
Spherical-cnn	98.4	85	46
s2cnn	96	90	17
R-SCNN (ours)	94	92	72

Citation

If you find the work useful, please cite as:

@ARTICLE{9713880,
  author={Mukhaimar, Ayman and Tennakoon, Ruwan and Lai, Chow Yin and Hoseinnezhad, Reza and Bab-Hadiashar, Alireza},
  journal={IEEE Access}, 
  title={Robust Object Classification Approach Using Spherical Harmonics}, 
  year={2022},
  volume={10},
  number={},
  pages={21541-21553},
  doi={10.1109/ACCESS.2022.3151350}}