Patch-Based Progressive 3D Point Set Upsampling

By Yifan Wang, Shihao Wu, Hui Huang, Daniel Cohen-Or, Olga Sorkine-Hornung. Computer Vision and Pattern Recognition, 2018

321 citations recorded by the source corpus.

We present a detail-driven deep neural network for point set upsampling. A high-resolution point set is essential for point-based rendering and surface reconstruction. Inspired by the recent success of neural image super-resolution techniques, we progressively train a cascade of patch-based upsampling networks on different levels of detail end-to-end. We propose a series of architectural design contributions that lead to a substantial performance boost. The effect of each technical contribution is demonstrated in an ablation study. Qualitative and quantitative experiments show that our method significantly outperforms the state-of-the-art learning-based and optimazation-based approaches, both in terms of handling low-resolution inputs and revealing high-fidelity details.

Subfields: Computational Mechanics, Computer Graphics and Computer-Aided Design, Computer Vision and Pattern Recognition

View Paper: https://www.semanticscholar.org/paper/c92c70dbfc2907350a2c547d6126da783aa5e1b6

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Computational Mechanics

1. P. Heng
2. Daniel Cohen-Or
3. Olga Sorkine-Hornung
4. Chi-Wing Fu
5. Ruihui Li
6. Shihao Wu
7. Yifan Wang
8. Philipp Krahenbuhl

In Computer Graphics and Computer-Aided Design

1. Daniel Cohen-Or
2. Olga Sorkine-Hornung
3. P. Heng
4. Chi-Wing Fu
5. Ruihui Li
6. Yifan Wang
7. Shihao Wu
8. Philipp Krahenbuhl

In Computer Vision and Pattern Recognition

1. Xingyi Zhou
2. P. Heng
3. Philipp Krahenbuhl
4. Tianwei Yin
5. Luc Van Gool
6. Daniel Cohen-Or
7. Chi-Wing Fu
8. Ming-Hsuan Yang

Citations in the corpus

• Differentiable surface splatting for point-based geometry processing (2019, methodology): “To synthesize training data for the Pix2Pix denoising network, we use the training set of the Sketchfab dataset [Yifan et al. 2018], which consist of 91 high-resolution 3D models.”
• Multimodal Virtual Point 3D Detection (2021, background): “[64] learn high level point-wise features from raw Lidar scans.”
• PU-GAN: A Point Cloud Upsampling Adversarial Network (2019, methodology): “Rather than a single-step expansion, the very recent MPU method [38] breaks a 16× upsampling network into four successive 2× upsampling subnets to progressively upsample in multiple steps.”
• Learning Gradient Fields for Shape Generation (2020, unknown): “[66] and consists of 2D contour points extracted from the MNIST [32] dataset, which contains 50K and 10K training and testing images.”