.png){kind=icon}
.png){kind=icon}
Return
Intermediate prototype network for few-shot segmentation
.png){kind=icon}
PRE
AI
DOI：10.1016/j.sigpro.2022.108811.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Few-shot segmentation aims to learn a model that can quickly adapt to new classes with limited labeled images. It remains challenging due to the large discrepancy of the targets between the support and query image, which hinders the label propagation from the support to query image. In this work, from a perspective of data distribution, we are committed to explicitly modeling an appropriate intermediate representation space to alleviate the target discrepancy between the support and query image. Specifically, we first propose the Intermediate Representation Module (IRM) that integrates the target information in the support and query image into an intermediate representation space, where the targets in the support and query image are represented as a Gaussian distribution respectively and the targets' distributional representations in the support and query image are probabilistically mixed to extend the target distribution. Second, we take one step further by incorporating an Intermediate Prototype Module (IPM) that replaces the target feature statistics in the support image with that in the intermediate representation space and then generates a new intermediate prototype to re-locate the target region in the query image to mitigate the target discrepancy. Extensive experiments on PASCAL-5(i), COCO- 20(i) and FSS-1000 demonstrate that our approach outperforms state-of-the-art methods. (c) 2022 Elsevier B.V. All rights reserved.
Keywords:
Few-shot segmentation
Few-shot learning
Semantic segmentation
Intermediate prototype