多重實例迴歸於定界框內物件輪廓之估測

Abstract

In this work, we address the high annotation cost of acquiring training data for semantic segmentation. Most modern approaches utilize graphical models, such as the conditional random fields, to carry out semantic segmentation, and hence rely on sufficient training data in form of object contours. To reduce the manual effort on annotating contours, we consider the training dataset for semantic segmentation is a mixture of a few object contours and an abundant set of bounding boxes of objects. Our idea is to borrow the knowledge derived from the object contours to infer the unknown object contours in the bounding boxes. The inferred contours then can serve as training data. To this end, we generate multiple contour hypotheses for each bounding box with the constraint that at least one hypothesis is close to the ground truth. Corresponding to multiple instance learning (MIL), a bounding box can be treated as a bag with its contour hypotheses as instances. We proposed an approach, called augmented multiple instance regression (AMIR), that formulates the task of hypothesis selection as the problem of MIR, and augments information derived from the object contours to guide and regularize the training process of MIR. The proposed approach is evaluated in the Pascal VOC segmentation task. The experimental results demonstrate that AMIR can precisely infer the object contours in the bounding boxes, and hence provides effective alternates of manually labeled contours for semantic segmentation.