Automatic Annotation of Object Instances by Region-Based Recurrent Neural Networks

In recent years, a wide variety of automatic, semiautomatic and manual approaches to image annotation have been proposed. These prerequisites have been driven by continuous advances of deep learning algorithms that often encounter the problem of insufficient or inappropriate training data, as well as sub-par markings’ accuracy which can have a direct impact on the model’s performance regardless. The main contribution of this paper is the development of a complex annotation framework able to automatically generate high-quality markings. The annotation work-flow aims to be an iterative process allowing automatic labeling of object bounding boxes, while simultaneously predicting the polygon outlining the object instance inside the box. The markings’ format is fully compatible with COCO Detection & Panoptic APIs that provide open-source interfaces for loading, parsing, and visualizing annotations. Following the completion of the research project funding this research, the code will be publicly available.