The key highlights and insights of the content are:
Existing deep learning methods for point cloud reconstruction and denoising rely on small datasets of 3D shapes. The authors circumvent this problem by leveraging deep learning methods trained on billions of images.
The authors propose a hybrid surface-appearance differentiable renderer that models normals and appearances using per-point spherical harmonics coefficients. This allows them to address shape reconstruction with changing lighting conditions.
To improve reconstruction in constraint settings, the authors introduce a semantic consistency regularization term that compares renderings of the point cloud from unseen camera poses with embeddings obtained from ground truth views.
The authors propose a diffusion-based network to denoise a wide variety of noise types from point cloud renderings. This is more robust to point cloud colors and lighting conditions compared to a GAN-based approach.
The authors show improved few-shot 3D shape reconstruction using semantic regularization and achieve similar quality to state-of-the-art methods while using less training images. They also demonstrate better point cloud denoising performance compared to a GAN-based network.
To Another Language
from source content
arxiv.org
Thông tin chi tiết chính được chắt lọc từ
by Pietro Bonaz... lúc arxiv.org 04-02-2024
https://arxiv.org/pdf/2404.01112.pdfYêu cầu sâu hơn