Kunho Kim

2022 Spring KAIST 3D ML Project

In this work, we reimplement and reproduce Score-Based Point Cloud Denoising (ICCV 2021).

Score-denoise [1] constructs the model that predicts the score of the distribution, and performs gradient ascent using the predicted score to denoise the point cloud. We use the same experimental setup as the original paper. We use PU-Net [2] dataset for training and testing our model, and PC-Net [3] dataset for testing only. We evaluate our model using two metrics: Chamfer distance (CD) [4] and point-to-mesh distance (P2M) [5]. Our model shows results that are almost similar to the author’s model, but slightly inferior. Especially for unsupervised learning, our model shows a slightly better result than the authors’ model. In summary, our achievements are:

• We implemented the whole code for original paper with a very basic skeleton code,
• We achieved almost similar but slightly inferior performance compared to the authors’ model.

Github, Report
    

Poster

References

[1] Luo Shitong and Hu Wei. Score-Based Point Cloud Denoising. ICCV 2021.
[2] Yu Lequan et al. PU-Net: Point Cloud Upsampling Network. CVPR 2018.
[3] Rakotosaona Marie-Julie et al. PointCleanNet: Learning to denoise and remove outliers from dense point clouds. Computer Graphics Forum, 2020.
[4] Fan Haoqiang et al. A Point Set Generation Network for 3D Object Reconstruction from a Single Image. arXiv 2016.
[5] Ravi Nikhila et al. Accelerating 3D Deep Learning with PyTorch3D. arXiv 2020.