Weakly supervised semantic segmentation is a challenging problem in computer vision filed. In this paper, we propose a novel method to perform weakly supervised semantic segmentation based on fast regional convolutional neural network (Faster R-CNN) and Grabcut. To deal with the imprecise location of Faster R-CNN detection, we provide a method to optimize the bounding box by building the undirected region adjacency graph on the image to search for the regions which belong to the object outside the bounding box. Then the obtained optimal bounding box is applied to initialize Grabcut algorithm. These accelerate the convergence of the Grabcut algorithm and no extra user interaction is needed. Our approach is validated on the PASCAL VOC 2012 and MSRC-21 datasets. The experimental results demonstrate that our method can provide efficient results in comparison with some state-of-the-arts.
Paper
Full text
Optimal bounding box and Grabcut for weakly supervised segmentation
Semantic Scholar · Computer Science · 2019
Abstract
Weakly supervised semantic segmentation is a challenging problem in computer vision filed. In this paper, we propose a novel method to perform weakly supervised semantic segmentation based on fast regional convolutional neural network (Faster R-CNN) and Grabcut. To deal with the imprecise location of Faster R-CNN detection, we provide a method to optimize the bounding box by building the undirected region adjacency graph on the image to search for the regions which belong to the object outside the bounding box. Then the obtained optimal bounding box is applied to initialize Grabcut algorithm. These accelerate the convergence of the Grabcut algorithm and no extra user interaction is needed. Our approach is validated on the PASCAL VOC 2012 and MSRC-21 datasets. The experimental results demonstrate that our method can provide efficient results in comparison with some state-of-the-arts.