Learning shape correspondence with anisotropic convolutional neural networks
Authors: Davide Boscaini, Jonathan Masci, Emanuele Rodolà, Michael Bronstein
NeurIPS 2016 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We tested ACNNs performance in challenging settings, achieving state-of-the-art results on recent correspondence benchmarks. In this section, we evaluate the proposed ACNN method and compare it to state-of-the-art approaches. In all experiments, we used L = 16 orientations and the anisotropy parameter = 100. For all experiments, training was done by minimizing the loss (10). Full mesh correspondence We used the FAUST humans dataset [3]. Partial correspondence We used the recent very challenging SHREC 16 Partial Correspondence benchmark [7]. |
| Researcher Affiliation | Collaboration | Davide Boscaini1, Jonathan Masci1, Emanuele Rodol a1, Michael Bronstein1,2,3 1USI Lugano, Switzerland 2Tel Aviv University, Israel 3Intel, Israel |
| Pseudocode | No | The paper does not contain structured pseudocode or algorithm blocks. |
| Open Source Code | No | The paper mentions that "Neural networks were implemented in Theano [2]" but does not provide a link or explicit statement about the availability of their own ACNN source code. |
| Open Datasets | Yes | We used the FAUST humans dataset [3], containing 100 meshes of 10 scanned subjects, each in 10 different poses. We used the recent very challenging SHREC 16 Partial Correspondence benchmark [7], consisting of nearly-isometrically deformed shapes from eight classes, with different parts removed. |
| Dataset Splits | Yes | First 80 shapes for training and the remaining 20 for testing, following verbatim the settings of [16]. The dataset was split into training and testing disjoint sets. For cuts, training was done on 15 shapes per class; for holes, training was done on 10 shapes per class. |
| Hardware Specification | Yes | For shapes with 6.9K vertices, Laplacian computation and eigendecomposition took 1 sec and 4 seconds per angle, respectively on a desktop workstation with 64Gb of RAM and i7-4820K CPU. |
| Software Dependencies | No | The paper states that "Neural networks were implemented in Theano [2]" but does not provide specific version numbers for Theano or any other software dependencies. |
| Experiment Setup | Yes | In all experiments, we used L = 16 orientations and the anisotropy parameter = 100. Neural networks were implemented in Theano [2]. The ADAM [11] stochastic optimization algorithm was used with initial learning rate of 10 3, β1 = 0.9, and β2 = 0.999. For this experiment, we adopted the following architecture inspired by GCNN [16]: FC64+IC64+IC128+IC256+FC1024+FC512+Softmax. The dropout regularization, with drop = 0.5, was crucial to avoid overfitting on such a small training set. We used the following ACNN architecture: IC32+FC1024+DO(0.5)+FC2048+DO(0.5)+Softmax. |