CoMasTRe: Continual Segmentation Framework with Objectness Learning and Class Recognition
Core Concepts
CoMasTRe disentangles segmentation into objectness learning and class recognition, improving performance in continual learning tasks.
Abstract
CoMasTRe proposes a two-stage segmenter for continual segmentation, focusing on objectness learning and class recognition. The method outperforms previous state-of-the-art models on PASCAL VOC 2012 and ADE20K datasets. By distilling objectness and classification knowledge separately, CoMasTRe achieves better stability and plasticity in learning new classes while preserving old class knowledge. The framework leverages the transfer ability of objectness to improve segmentation performance. Extensive experiments demonstrate the effectiveness of CoMasTRe in tackling forgetting issues in continual semantic segmentation.
Continual Segmentation with Disentangled Objectness Learning and Class Recognition
Stats
CoMasTRe reaches a new state-of-the-art on both PASCAL VOC 2012 and ADE20K datasets.
The method outperforms per-pixel methods on new classes by up to 32.16% on PASCAL VOC.
CoMasTRe significantly outperforms prior arts in all classes on ADE20K.
Quotes
"Most continual segmentation methods tackle the problem as a per-pixel classification task."
"Based on these findings, we propose CoMasTRe by disentangling continual segmentation into two stages: forgetting-resistant continual objectness learning and well-researched continual classification."
"We assess the effectiveness of CoMasTRe on PASCAL VOC and ADE20K."