[3Rin4] Interactive 1

[3Rin4-55] Efficient Construction of training data for Object Detection

〇Hirotaka Tanaka¹, Hiroyuki Shinnou¹ (1.Ibaraki University)

Keywords:semantic segmentation

Object detection is one of the image recognition tasks and generally be solved by supervised learning. The supervised data for object detection is composed of a bounding box and a class label. Construction of such supervised data is generally expensive because it is performed manually. The purpose of this study is to construct an accurate bounding box for supervised data. Although it is costly to construct an accurate bounding box by hand, a bounding box that does not require accuracy can be constructed at relatively low cost even by hand. Therefore, we propose a method for construction of the accurate bounding box from a low-cost bounding box that has room for objects. In the proposed method, the object region is estimated by semantic segmentation for the region inside the bounding box constructed by hand, and the box surrounding the entire estimated region is an accurate bounding box. With this method, more accurate bounding box could be constructed from a bounding box that has room for the object.