Segmentation Mask Refinement Using Image Transformations

Abstract

This paper discusses object proposal generation, which is a crucial step of instance-level semantic segmentation (instance segmentation). Known as a challenging computer vision task, the instance segmentation requires jointly detecting and segmenting individual instances of objects in an image. A common approach to this task is first to propose a set of class-agnostic object candidates in the forms of segmentation masks, which represent both object locations and boundaries, and then to perform classification on each object candidate. In this paper, we propose an effective refinement process that employs image transformations and mask matching to increase the accuracy of object segmentation masks. The proposed refinement process is applied to three state-of-the-art object proposal methods (DeepMask, SharpMask, and FastMask), and is evaluated on two standard benchmarks (Microsoft COCO and PASCAL VOC). Both the quantitative and qualitative results show the effectiveness of the process across various experimental settings.