The Effect of Improving Annotation Quality on Object Detection Datasets: A Preliminary Study

Jiaxin Ma, Yoshitaka Ushiku, Miori Sagara

Abstract: In this study, we partially reannotate conventional benchmark datasets for object detection and check whether there is performance improvement/drop compared with the original annotations. Recent studies on the annotation qualities of ImageNet for image classification revealed some issues of how to associate only a single label to each image accurately. Object detection, on the other hand, should have other nontrivial issues because there are multiple objects in a single image, and realizing consistency among bounding boxes is challenging. A team of professional annotators was formed for MS COCO and Google Open Images datasets. To realize highly-consistent an-notations, we prepared carefully designed guidelines for each category and selected quality inspectors who checked the annotation quality of each annotator. Finally, we applied conventional object detection methods for reannotated parts of each dataset. We found mixed results: whether the performance dropped or improved depended on each category and dataset.