Research On Sampling Strategies For Implicit Neural Representation In 3D Reconstruction

In recent years,with the widespread application of deep learning technology,high-quality and low-cost 3D reconstruction techniques have experienced rapid development,and researchers promote the progress of the 3D reconstruction field from various directions.Among them,the implicit neural representation,which implicitly represents the object shape as the decision boundary in 3D space,has become the focus of interest for an expanding group of researchers,due to its ability to represent continuous shapes and generate 3D objects at arbitrary resolutions.For a query point at any position in space,the implicit neural representation can predict the spatial property of the point and extract the 3D shape by finding the decision boundary of the properties of different points.Therefore,implicit neural representation requires a certain number of query points as learning samples in the training process of the neural network.However,different implicit neural representation networks exhibit significant differences in the density and spatial distribution of query points when sampling points on 3D objects.The difference in sampling strategies of query points affects the network performance,and some researchers have noted the importance of sampling strategies but have only limited it to simple comparisons based on the author’s current model,without conducting more in-depth research.This paper focuses on the sampling strategy of query points in the training process of implicit neural representation networks.From a macro perspective in relevant fields,a systematic analysis research is conducted on how sampling strategies affect network performance and how to choose the correct sampling strategy for the network.The main research contents are as follows:(1)The Impact of Sampling Strategies on Implicit Neural Representation.It is observed that various implicit neural representation networks have different preferences for sampling strategies of query points,and we consider this from two angles: density and spatial distribution.Firstly,regarding the sampling density strategy of query points,we set two modes: sparse sampling and dense sampling,and conducted experiments on different networks.We verified the unnecessary nature of dense sampling and showed that sparse sampling is sufficient to balance model performance and experimental cost.Secondly,we conducted in-depth research on the relationship between sampling distribution strategies and network types.By analyzing the differences in network structure,we classified the relevant networks into three types,then demonstrated and analyzed through comparative experiments the rationality of network classification and the impact of network type differences on the choice of sampling distribution strategies.We also explained the situation where two network types are incompatible with sampling distribution strategies.Finally,we discussed the relationship between sampling distribution strategies and implicit functions and concluded that classification implicit functions have higher fault tolerance rates and are more suitable as implicit neural representation methods.(2)The Improvement of Sampling Strategies for Implicit Neural Representation.We propose two improvements to existing sampling strategies for implicit neural representations.First,we introduce a sampling strategy based on linear distribution that gradually transitions the sample distribution from near-surface to far-surface positions by adjusting the spatial distribution of query points.Our proposed linear sampling strategy is more robust and performs well on various models.Second,we find that errors in reconstruction often occur at far-surface positions,so we propose a sampling strategy based on distance mask that masks out far-surface query points which outside of a certain distance range to prevent the network from making erroneous predictions.The distance mask strategy ensures compatibility between the network and sampling strategy,and enables the neural network to focus more on fine surface reconstruction.