Dual Octree Graph Networks for Learning Adaptive Volumetric Shape Representations

Proceedings Paper Computer Science, Artificial Intelligence

Neural Fields as Learnable Kernels for 3D Reconstruction

Francis Williams et al.

Summary: This paper presents Neural Kernel Fields, a new method for reconstructing implicit 3D shapes based on a learned kernel ridge regression. The method achieves state-of-the-art results in reconstructing 3D objects and large scenes from sparse oriented points, and also demonstrates strong generalization capability to shape categories outside the training set. The key insight of the approach is that kernel methods are highly effective for shape reconstruction when an appropriate inductive bias is chosen.

2022 IEEE/CVF CONFERENCE ON COMPUTER VISION AND PATTERN RECOGNITION (CVPR 2022) (2022)

Add to Collection

Article Computer Science, Software Engineering

Acorn: Adaptive Coordinate Networks for Neural Scene Representation

Julien N. P. Martel et al.

Summary: Neural representations have emerged as a new paradigm in various applications, offering flexibility and accuracy at moderate resolutions but facing challenges in representing large-scale or complex scenes. A new hybrid implicit-explicit network architecture and training strategy has been introduced, showing promising results in fitting high-resolution images and 3D shapes while significantly reducing training times and memory requirements.

ACM TRANSACTIONS ON GRAPHICS (2021)

Add to Collection

Proceedings Paper Computer Science, Artificial Intelligence

Adaptive Surface Reconstruction with Multiscale Convolutional Kernels

Benjamin Ummenhofer et al.

Summary: The study introduces generalized convolutional kernels for 3D reconstruction from point clouds, utilizing multiscale convolutional kernels that can be applied to adaptive grids and able to predict signed and unsigned distance fields in large datasets.

2021 IEEE/CVF INTERNATIONAL CONFERENCE ON COMPUTER VISION (ICCV 2021) (2021)

Add to Collection

Proceedings Paper Computer Science, Artificial Intelligence

Neural Geometric Level of Detail: Real-time Rendering with Implicit 3D Shapes

Towaki Takikawa et al.

Summary: The study introduces an efficient neural representation that enables real-time rendering of high-fidelity neural SDFs and achieves state-of-the-art geometry reconstruction quality.

2021 IEEE/CVF CONFERENCE ON COMPUTER VISION AND PATTERN RECOGNITION, CVPR 2021 (2021)

Add to Collection

Proceedings Paper Computer Science, Artificial Intelligence

Neural Splines: Fitting 3D Surfaces with Infinitely-Wide Neural Networks

Francis Williams et al.

Summary: Neural Splines is a 3D surface reconstruction technique based on random feature kernels from infinitely-wide shallow ReLU networks, achieving state-of-the-art results. The method is easy to analyze and can be accelerated, serving as a generalization of cubic spline interpolation.

2021 IEEE/CVF CONFERENCE ON COMPUTER VISION AND PATTERN RECOGNITION, CVPR 2021 (2021)

Add to Collection

Article Computer Science, Software Engineering