FAST ROTATED BOUNDING BOX ANNOTATIONS FOR OBJECT DETECTION

Abstract

Traditionally, object detection models use a large amount of annotated data and axis-aligned bounding boxes (AABBs) are often chosen as the image annotation technique for both training and predictions. The purpose of annotating the objects in the images is to indicate the regions of interest with the corresponding labels. Accurate object annotations help the computer vision models to understand the distinct patterns of the image features to recognize and localize different classes of objects. However, AABBs are often a poor fit for elongated object instances. It’s also challenging to localize objects with AABBs in densely packed aerial images because of overlapping adjacent bounding boxes. Alternatively, using rectangular annotations that can be oriented diagonally, also known as rotated bounding boxes (RBB), can provide a much tighter fit for elongated objects and reduce the potential bounding box overlap between adjacent objects. However, RBBs are much more time-consuming and tedious to annotate than AABBs for large datasets.

In this work, we propose a novel annotation tool named as FastRoLabelImg (Fast Rotated LabelImg) for producing high-quality RBB annotations with low time and effort. The tool generates accurate RBB proposals for objects of interest as the annotator makes progress through the dataset. It can also adapt available AABBs to generate RBB proposals. Furthermore, a multipoint box drawing system is provided to reduce manual RBB annotation time compared to the existing methods. Across three diverse datasets, we show that the proposal generation methods can achieve a maximum of 88.9% manual workload reduction. We also show that our proposed manual annotation method is twice as fast as the existing system with the same accuracy by conducting a participant study. Lastly, we publish the RBB annotations for two public datasets in order to motivate future research that will contribute in developing more competent object detection algorithms capable of RBB predictions.