Published in Conference October 16-19, 2022
Published at the 2022 IEEE International Conference on Image Processing, in October 2022, the article “Conditional Reconstruction for Open-Set Semantic Segmentation” deals with a new method for modeling the segmentation of open sets. Called CoReSeg, the new technique addresses the problem using class-conditioned reconstruction of the input images according to their pixel-by-pixel mask. According to the authors, open set segmentation is a relatively new and unexplored task, with only a few methods proposed to model such tasks. The new proposed methodology conditions each input pixel to all known classes, expecting larger errors for pixels from unknown classes.
According to the authors, it was observed that the proposed method produces better semantic consistency in its predictions than the baselines, resulting in cleaner segmentation maps that better adjust to the object boundaries. “CoReSeg outperforms state-of-the-art methods on the Vaihingen and Potsdam ISPRS datasets, while being competitive on the Houston 2018 IEEE GRSS Data Fusion dataset,” they state.
The new official implementation for CoReSeg is available at: https://github.com/iannunes/CoReSeg. Additionally, there is a virtual demonstration of Colab for CoReSeg in Vaihingen buildings, with page visitors invited to test the script on Google Colab: https://drive.google.com/drive/folders/1tBcvLolYpMCc8-EA-9OO5IfBU2eL1Cff?usp =sharing,
The article is authored by researchers Ian Nunes, from the Pontifical Catholic University of Rio de Janeiro, Brazil; Matheus B. Pereira, from the Federal University of Minas Gerais, Belo Horizonte, Brazil; Hugo Oliveira, from the University of São Paulo, Brazil; Jefersson A. dos Santos, Federal University of Minas Gerais, Belo Horizonte, Brazil and University of Stirling, Stirling, Scotland, United Kingdom; and Marcus Poggi, from the Pontifical Catholic University of Rio de Janeiro, Brazil.