Gauge Equivariant Convolutional Networks and the Icosahedral CNN

Authors: Taco Cohen, Maurice Weiler, Berkay Kicanaoglu, Max Welling

ICML 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Results on Ico MNIST, climate pattern segmentation, and omnidirectional RGB-D image segmentation are presented in Sec. 5.
Researcher Affiliation Collaboration 1Qualcomm AI Research, Amsterdam, NL. 2Qualcomm-University of Amsterdam (QUVA) Lab.
Pseudocode No The paper describes algorithms but does not provide structured pseudocode blocks or algorithms labeled as such.
Open Source Code No The paper does not provide a statement or link to open-source code for the described methodology.
Open Datasets Yes We use the exact same data and evaluation methodology as (Jiang et al., 2018). The preprocessed data as released by (Jiang et al., 2018) consists of 16-channel spherical images at resolution r = 5, which we reinterpret as icosahedral signals (introducing slight distortion). See (Mudigonda et al., 2017) for a detailed description of the data.
Dataset Splits Yes We generate three different versions of the training and test sets, differing in the transformations applied to the data.
Hardware Specification No The paper mentions experiments were run but does not specify any hardware details such as GPU/CPU models or types.
Software Dependencies No The paper mentions 'conv2d' and deep learning primitives but does not specify any software dependencies with version numbers.
Experiment Setup Yes Our main model consists of one gauge equivariant scalar-to-regular (S2R) convolution layer, followed by 6 regular-to-regular (R2R) layers and 3 FC layers (see Supp. Mat. for architectural details). We also evaluate a model that uses only S2R convolution layers, followed by orientation pooling (a max over the 6 orientation channels of each regular feature, thus mapping a regular field to a scalar), as in (Masci et al., 2015). Finally, we consider a model that uses only rotation-invariant filters, i.e. scalar-to-scalar (S2S) convolutions, similar to standard graph CNNs (Boscaini et al., 2015; Kipf & Welling, 2017). We also compare to the fully SO(3)-equivariant but computationally costly Spherical CNN (S2CNN). See supp. mat. for architectural details and computational complexity analysis.