The research paper by Zhong, Bhattad, Wang, and Forsyth from the University of Illinois Urbana-Champaign investigates the lack of equivariance in state-of-the-art depth and normal predictors. Despite strong performances, these predictors do not respect the equivariant property to cropping-and-resizing. The researchers propose an equivariant regularization technique to promote cropping-and-resizing equivariance in depth and normal networks. The approach is applicable to both CNN and Transformer architectures and improves the supervised and semi-supervised learning performance of dense predictors on Taskonomy tasks. Finetuning with their loss on unlabeled images improves not only equivariance but also accuracy.
Publication date: 28 Sep 2023
Project Page: GitHub link not provided
Paper: https://arxiv.org/pdf/2309.16646
