Click here to Download the project abstract/base paper of the Fourier neural operator of the 3d reconstruction final year project.
WeintroduceNeuralPoissonSurfaceReconstruction(nPSR), anarchitectureforshapereconstructionthataddressesthe challengeofrecovering3Dshapesfrompoints.Traditional deep neural networks face challenges with common 3D-shaped cretization techniques due to their computational complexity at higher resolutions. To overcome this, we leverage Fourier NeuralOperators to solve the Poisson equation and construct a mesh from oriented point cloud measurements.nPSRexhibitstwomainadvantages: First, it enables efficient training on low-resolution data while achieving comparable performance at high-resolution evaluation,thankstotheresolution-agnosticnatureofFNOs. Thisfeatureallowsforone-shotsuper-resolution. Second, our method construction construction points sampling rates. Overall, euralPoissonsurface reconstruction not only improves upon the limitations of classical deep neural networks inspire construction but also achieves superior results in reconstruction quality, running time, and resolution agnosticism.
Introduction to 3d reconstruction, the process of creating a three-dimensional mesh of an object from 2D images or point clouds plays a crucial role in computer vision. Itisanactive research is with significant applications in robotics, virtual reality, and autonomous driving. A common sub in this field is the task of reconstructing the surface from noisy point clouds, which are typically acquired using techniques such as photogrammetry (multiple images) or depth sensors.
A plethora of shape reconstruction algorithms have been developed, and recently, Advanced deep learning approaches trainedonlargedatasetsof3Dobjects, such as ConvolutionalOccupancyNetworks [16]andPOCO[2], have emerged, promising improved reconstruction quality. However, the methods typically do not generalise distribution Another challenge faced by deep learning approaches is scalability, especially when dealing with complex models at high resolutions. Convolutional neural networks have achieved widespread success in computer vision and can straightforwardly generalise with dimensions. However, they operate on regular grids which limits their ability to represent fine details accurately. To overcome this limitation, various strategies have been proposed. have emerged, promising improved reconstruction quality. However, these methods typically do not generalise beyond the training distribution Another challenge faced by deep learning approaches is scalability, especially when dealing with complex models at high resolutions.
