Interactive Image Segmentation based on Laplace Transform
Xinping Cheng, Yuxiang Qian, Tao Wang · 2025
Interactive image segmentation aims to extract target object of interest to the user with high precision under the guidance of user interaction. SimpleClick ${ }^{[1]}$ introduced the Transformer architecture for the first time to interactive segmentation and significantly improved segmentation performance. However, as Liu ${ }^{[1]}$ elaborated, SimpleClick may fail with small targets due to low confidence. We believe that the problem of low confidence is due to class imbalance caused by softmax calculation. We have designed a Laplacian Transformer that enhances the discrimination between foreground and background and increases confidence for thin objects without adding additional calculations and parameters. Vast experiments on the popular GrabCut, Berkeley, DAVIS, SBD, Pascal VOC datasets verified the effectiveness of the proposed method. Remarkably, on the large-scale datasets SBD and DAVIS, our model achieves 5.23 in NoC@90 and 4.78 in NoC@90 respectively, outperforming SimpleClick 0.39 and 0.28.