Core Concepts
提案されたCoBraは、CNNとViTの補完的な知識を融合し、強力な疑似マスクを生成する新しいデュアルブランチフレームワークです。
Abstract
Abstract:
CoBra proposes a dual branch framework to fuse CNN and ViT knowledge.
Extensive experiments show state-of-the-art results on PASCAL VOC 2012 and MS COCO 2014 datasets.
Introduction:
Weakly supervised semantic segmentation aims to leverage image-level class labels.
Prior works focused on utilizing Class Activation Maps (CAMs) for object localization.
Methods:
CoBra consists of Class-Aware Knowledge (CAK) and Semantic-Aware Knowledge (SAK) branches.
CAP and SAP are used to exchange complementary knowledge between branches.
Results:
CoBra achieves the best seed and mask performance compared to existing methods.
State-of-the-art results are obtained on PASCAL VOC 2012 dataset for both ResNet101 and MiT-B2 backbones.
Conclusion:
CoBra demonstrates the importance of exchanging class and semantic knowledge in weakly supervised semantic segmentation.
Stats
提案されたCoBraは、PASCAL VOC 2012およびMS COCO 2014データセットで最先端のWSSS結果を示しています。
CNN CAMsとViT CAMsに関する様々な損失関数の影響も詳細に調査されました。