[论文解读] FPNN: Field Probing Neural Networks for 3D Data

Yangyan Li, Sören Pirk+3 more|arXiv (Cornell University)|May 20, 2016

3D Shape Modeling and Analysis参考文献 33被引用 209

一句话总结

该论文提出 Field Probing Neural Networks (FPNN)，它们在学习探针点的位置和权重以高效从3D场域提取特征，同时在3D对象分类上实现了与标准3D CNNs竞争力但计算量更低。

ABSTRACT

Building discriminative representations for 3D data has been an important task in computer graphics and computer vision research. Convolutional Neural Networks (CNNs) have shown to operate on 2D images with great success for a variety of tasks. Lifting convolution operators to 3D (3DCNNs) seems like a plausible and promising next step. Unfortunately, the computational complexity of 3D CNNs grows cubically with respect to voxel resolution. Moreover, since most 3D geometry representations are boundary based, occupied regions do not increase proportionately with the size of the discretization, resulting in wasted computation. In this work, we represent 3D spaces as volumetric fields, and propose a novel design that employs field probing filters to efficiently extract features from them. Each field probing filter is a set of probing points --- sensors that perceive the space. Our learning algorithm optimizes not only the weights associated with the probing points, but also their locations, which deforms the shape of the probing filters and adaptively distributes them in 3D space. The optimized probing points sense the 3D space "intelligently", rather than operating blindly over the entire domain. We show that field probing is significantly more efficient than 3DCNNs, while providing state-of-the-art performance, on classification tasks for 3D object recognition benchmark datasets.

研究动机与目标

• Motivate efficient discriminative feature learning for 3D data beyond cubic-cost 3D CNNs.
• Propose a field probing framework where filters are defined by trainable probing points and weights.
• Demonstrate that learning both probe locations and weights yields long-range, sparse-sensing capabilities.
• Show that field probing layers provide competitive accuracy with significantly reduced computation on 3D classification benchmarks.

提出的方法

• Represent 3D data as volumetric fields (e.g., distance fields, normal fields).
• Replace conventional 3D convolutions with field probing layers consisting of Sensor, DotProduct, and Gaussian layers.
• Train both probing point locations and filter weights via backpropagation.
• Use a Gaussian transform on distance field values to emphasize surface-adjacent samples.
• Initialize probing points broadly and let learning adjust their positions to sense informative regions.

实验结果

研究问题

• RQ1 Can field probing filters learn to sense 3D space efficiently by optimizing both probe locations and weights?
• RQ2 Do field probing layers achieve competitive 3D object classification accuracy with lower computational cost than 3D CNNs across varying resolutions and sparsity levels?
• RQ3 Are the learned features robust to spatial perturbations and transferable across datasets?
• RQ4 How does increasing input field resolution and incorporating multiple fields affect performance?

主要发现

• Field probing layers provide substantial accuracy gains over baseline (e.g., 1-FC setting improves from 79.1% to 85.0%).
• A deeper 4-FC network with field probing yields 87.5% accuracy, with a smaller gap to the baseline than the shallower network.
• Using multiple input fields (distance and normals) yields consistent performance gains.
• The method shows robustness to spatial perturbations (rotations, translations, and scaling).
• Higher input field resolutions improve performance while computational cost remains largely independent of resolution.