Code underlying the publication: Color Equivariant Convolutional Networks

doi: 10.4121/089a228a-bd6c-487b-98c6-3302a39b3108.v1
The doi above is for this specific version of this dataset, which is currently the latest. Newer versions may be published in the future. For a link that will always point to the latest version, please use
doi: 10.4121/089a228a-bd6c-487b-98c6-3302a39b3108
Datacite citation style:
Attila Lengyel; van Gemert, Jan; Bruintjes, Robert-Jan; Strafforello, Ombretta; Gielisse, Alexander (2023): Code underlying the publication: Color Equivariant Convolutional Networks. Version 1. 4TU.ResearchData. software. https://doi.org/10.4121/089a228a-bd6c-487b-98c6-3302a39b3108.v1
Other citation styles (APA, Harvard, MLA, Vancouver, Chicago, IEEE) available at Datacite
Software

Code corresponding to NeurIPS 2023 conference paper "Color Equivariant Convolutional Networks".


Abstract

Color is a crucial visual cue readily exploited by Convolutional Neural Networks (CNNs) for object recognition. However, CNNs struggle if there is data imbalance between color variations introduced by accidental recording conditions. Color invariance addresses this issue but does so at the cost of removing all color information, which sacrifices discriminative power. In this paper, we propose Color Equivariant Convolutions (CEConvs), a novel deep learning building block that enables shape feature sharing across the color spectrum while retaining important color information. We extend the notion of equivariance from geometric to photometric transformations by incorporating parameter sharing over hue-shifts in a neural network. We demonstrate the benefits of CEConvs in terms of downstream performance to various tasks and improved robustness to color changes, including train-test distribution shifts. Our approach can be seamlessly integrated into existing architectures, such as ResNets, and offers a promising solution for addressing color-based domain shifts in CNNs.

history
  • 2023-11-29 first online, published, posted
publisher
4TU.ResearchData
funding
  • Tabula Inscripta: Prior knowledge for deep learning (grant code VI.Vidi.192.100) [more info...] Dutch Research Council
organizations
TU Delft, TU Delft, Faculty of Electrical Engineering, Mathematics and Computer Science, Department of Intelligent Systems, Computer Vision Lab

DATA

To access the source code, use the following command:

git clone https://data.4tu.nl/v3/datasets/2ab28fae-8cbf-4640-8da3-c284e89dc06c.git