Introducing a self-supervised, superfeature-based network for video object segmentation

Título: Introducing a self-supervised, superfeature-based network for video object segmentation

Autores: Marcelo Mendonça, Luciano Oliveira

Resumo: This 3-page long paper summarizes our PhD thesis with the aim of participating in the CBIC23 Contest of Theses and Dissertations (CTD). Our thesis introduces a novel videoobject segmentation (VOS) method, called SHLS, that uses superpixels to build a high-compressed latent space. The proposed method is completely self-supervised, initially trained on a dataset of various orders of magnitude less than existing self-supervised VOS methods; using pseudo-labels in the offline training stage avoids the burden of annotations. The ultra-compact latent space allows for creating more efficient memory clusters, ultimately speeding up the segmentation process across the video. With efficient-oriented memory usage, SHLS achieved superior performance on single-object segmentation and comparable results with other state-of-the-art methods on multi-object segmentation on the DAVIS dataset.

Palavras-chave: component, formatting, style, styling, insert

Páginas: 4

Artigo em pdf: CBIC_2023_paper_CTDD03.pdf