Publicação
General unsupervised semantic segmentation pipeline
| Resumo: | In this work, we present a Convolutional Neural Network-based unsupervised semantic segmentation pipeline, with a focus on remote sensing imagery. Our pipeline improves upon the current state-of-art papers in literature, resulting in a versatile methodology that can take supervised, unsupervised, and weakly supervised inputs. We also propose an automatic scribble generation methodology that is able to semi-automatically label large datasets given minimal supervision. To accompany this methodology we also propose a scribble classifier and a scribble labeling tool. And finally, we propose two real-world applications where we test the capabilities of our proposed network. Results on benchmarking datasets show that our proposed network can be competitive with the current state of the art for CNN-based methods, the scribble generator is able to provide meaningful and relevant scribbles for large datasets, and our applications show a variety of possibilities of usage for our network. |
|---|---|
| Autores principais: | João Pedro Klock Ferreira |
| Assunto: | Semantic segmentation Unsupervised Weakly supervised Remote sensing Scribbles Engenharia elétrica Redes Neurais (Computação) Sensoriamento remoto |
| Ano: | 2023 |
| País: | Brasil |
| Tipo de documento: | dissertação de mestrado |
| Tipo de acesso: | unknown |
| Instituição associada: | IBICT |
| Idioma: | inglês |
| Origem: | Oasisbr |
| Resumo: | In this work, we present a Convolutional Neural Network-based unsupervised semantic segmentation pipeline, with a focus on remote sensing imagery. Our pipeline improves upon the current state-of-art papers in literature, resulting in a versatile methodology that can take supervised, unsupervised, and weakly supervised inputs. We also propose an automatic scribble generation methodology that is able to semi-automatically label large datasets given minimal supervision. To accompany this methodology we also propose a scribble classifier and a scribble labeling tool. And finally, we propose two real-world applications where we test the capabilities of our proposed network. Results on benchmarking datasets show that our proposed network can be competitive with the current state of the art for CNN-based methods, the scribble generator is able to provide meaningful and relevant scribbles for large datasets, and our applications show a variety of possibilities of usage for our network. |
|---|
Atividades financiadas
Não foram encontrados projetos financiados.