FPNN: Field Probing Neural Networks for 3D Data

Authors: Yangyan Li, Soeren Pirk, Hao Su, Charles R. Qi, Leonidas J. Guibas

NeurIPS 2016 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We evaluate our field probing based neural networks (FPNN) on a classification task on Model Net [31] dataset, and show that they match the performance of 3DCNNs while requiring much less computation, as they are designed and trained to respect the sparsity of 3D data. (...) 4 Results and Discussions
Researcher Affiliation Academia Yangyan Li1,2 Sören Pirk1 Hao Su1 Charles R. Qi1 Leonidas J. Guibas1 (...) 1Stanford University, USA 2Shandong University, China
Pseudocode No The paper describes the architecture and components of the Field Probing Neural Network, but it does not include any explicit pseudocode or algorithm blocks.
Open Source Code Yes We open-source our code at https://github.com/yangyanli/FPNN for encouraging future developments.
Open Datasets Yes We use Model Net40 [31] (12,311 models from 40 categories, training/testing split with 9,843/2,468 models4) the standard benchmark for 3D object classification task, in our experiments.
Dataset Splits No The paper states 'training/testing split with 9,843/2,468 models' for ModelNet40, but it does not explicitly provide details for a separate validation split.
Hardware Specification Yes Figure 6: Running time of convolutional layers (same settings as that in [31]) and field probing layers (C N T = 1024 8 4) on Nvidia GTX TITAN with batch size 83.
Software Dependencies No The paper states 'We implemented our field probing layers in Caffe [12].' but does not specify a version number for Caffe or any other software dependencies with their versions.
Experiment Setup Yes We train our FPNN 80, 000 iterations on 64 64 64 distance field with batch size 1024.5, with SGD solver, learning rate 0.01, momentum 0.9, and weight decay 0.0005. (...) The σ hyper-parameter in Gaussian layer controls how sharp is the transform. We select its value empirically in our experiments, and the best performance is given when we use σ 10% of the object size.