.png){kind=icon}
.png){kind=icon}
Return
DiffClick: Click-differentiated enhancement network for interactive segmentation
.png){kind=icon}
PRE
AI
DOI：10.1016/j.patcog.2025.112217.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Click-based interactive segmentation aims to achieve precise segmentation using minimal positive and negative clicks. Existing methods often overlook the differences between positive and negative clicks. They have different objectives, and the number of negative clicks is far less than the number of positive ones. This leads to inadequate background refinement in the absence of negative clicks. In response, we propose DiffClick, a novel framework that processes positive and negative clicks in different enhancement modules. Starting from the initial segmentation results derived from the first click, both our Foreground Enhancement Module and Background Enhancement Module utilize a weight fusion module to augment features based on the type of guidance received. The Foreground Enhancement Module refines foreground features guided by positive clicks. Similarly, the Background Enhancement Module processes negative clicks and improves background segmentation by incorporating weight maps of background regions. Additionally, this module includes a non-target prototype to provide supplementary background guidance, ensuring effective segmentation even when negative clicks are lacking. Extensive experiments show that DiffClick beats most existing methods, especially when negative clicks are absent.
Keywords:
click-based interactive segmentation
foreground enhancement
background refinement
weight fusion module
non-target prototype
Journal
IF:
7.6
Papers:
1.3W
Citations:
4.5W