Automatic bounding-box-labeling method of occluded objects in virtual image data

Abstract

Computer vision technology is widely used based on its massive and correct data set, of which the bounding box labeling is a common method. Aimed at a large number of original image data set produced by virtual simulation, we proposed an automatic pixel-level bounding-box-labeling method to solve problem of accuracy and speed. The method starts by a fundamental algorithm based on targeted bounding box, which will be adopted to label the images produced by virtual simulation and learn from the bounding box of different objects; Next, the method will find consistent seed points and apply region growing algorithm to automatically produce binary images based on the seed points; Then, an occlusion-estimating algorithm can be used to evaluate the occluded conditions in the binary image; Finally, employ bounding-box-labeling algorithm to label targeted objects according to various occlusion. Apply the data set from 2019 Small Target Competition held by China Society of Images and Graphics to test and verify our method, the result turns out that this method can solve the occlusion problem especially the truncate occlusion and can label the objects' entire body precisely.