Título: | OPEN-SET SEMANTIC SEGMENTATION FOR REMOTE SENSING IMAGES | ||||||||||||
Autor: |
IAN MONTEIRO NUNES |
||||||||||||
Colaborador(es): |
MARCUS VINICIUS SOLEDADE POGGI DE ARAGAO - Orientador HUGO NEVES DE OLIVEIRA - Coorientador |
||||||||||||
Catalogação: | 21/MAR/2023 | Língua(s): | ENGLISH - UNITED STATES |
||||||||||
Tipo: | TEXT | Subtipo: | THESIS | ||||||||||
Notas: |
[pt] Todos os dados constantes dos documentos são de inteira responsabilidade de seus autores. Os dados utilizados nas descrições dos documentos estão em conformidade com os sistemas da administração da PUC-Rio. [en] All data contained in the documents are the sole responsibility of the authors. The data used in the descriptions of the documents are in conformity with the systems of the administration of PUC-Rio. |
||||||||||||
Referência(s): |
[pt] https://www.maxwell.vrac.puc-rio.br/projetosEspeciais/ETDs/consultas/conteudo.php?strSecao=resultado&nrSeq=62040&idi=1 [en] https://www.maxwell.vrac.puc-rio.br/projetosEspeciais/ETDs/consultas/conteudo.php?strSecao=resultado&nrSeq=62040&idi=2 |
||||||||||||
DOI: | https://doi.org/10.17771/PUCRio.acad.62040 | ||||||||||||
Resumo: | |||||||||||||
Collecting samples that exhaust all possible classes for real-world tasks is
usually difficult or impossible due to many different factors. In a realistic/feasible scenario, methods should be aware that the training data is incomplete and
that not all knowledge is available. Therefore all developed methods should be
able to identify the unknown samples while correctly executing the proposed
task to the known classes in the tests phase.
Open-Set Recognition and Semantic Segmentation models emerge to
handle this kind of scenario for, respectively, visual recognition and dense
labeling tasks. Initially, this work proposes a novel taxonomy aiming to
organize the literature and provide an understanding of the theoretical trends
that guided the existing approaches that may influence future methods. This
work tested the proposed techniques on remote sensing data, establishing new
state-of-the-art results for the used datasets. Remote sensing data differs from
RGB data as it deals with a plethora of sensors and with a high geographical
variation.
Open set segmentation is a relatively new and unexplored task, with
just a handful of methods proposed to model such tasks. This work also
proposes two distinct techniques to perform open-set semantic segmentation.
First, a method called OpenGMM extends the OpenPCS framework using a
Gaussian Mixture of Models to model the distribution of pixels for each class
in a multimodal manner. Second, the Conditional Reconstruction for Openset Semantic Segmentation (CoReSeg) method tackles the issue using classconditioned reconstruction of the input images according to their pixel-wise
mask. CoReSeg conditions each input pixel to all known classes, expecting
higher errors for pixels of unknown classes.
Qualitative results observation suggested that both proposed methods
produce better semantic consistency in their predictions than the baselines,
resulting in cleaner segmentation maps that better fit object boundaries. Also,
OpenGMM and CoReSeg outperformed state-of-the-art baseline methods on
Vaihingen and Potsdam ISPRS datasets.
The third proposed approach is a general post-processing procedure that
uses superpixels to enforce highly homogeneous regions to behave equally,
rectifying erroneous classified pixels within these regions. We also proposed
a novel superpixel generation method called FuSC.
All proposed approaches improved the quantitative and the qualitative
results for both datasets. Besides that, CoReSeg s prediction post-processed
with FuSC achieved state-of-the-art results for both datasets.
The official implementation of all proposed approaches is available at
https://github.com/iannunes.
|
|||||||||||||
|