Expand Object Detection Dataset Automatically via Deep Learning Neural Networks

Object detection is an important research field in image processing. The state-of-the-art machine learning models are trained with huge number of samples to solve this problem. Manually labeling objects on images is labor intensive, expensive, and time consuming. It is desirable to create an automatic labeling system to reduce the cost and the time needed. A deep learning neural networks based automatic labeling system is proposed in this paper to expand Microsoft Common Objects in Context 2017 (COCO 2017) dataset from 80 classes to 81 classes by adding labels for the new class, human faces, automatically. The system contains an Auto-Labeling Model (ALM) which is trained using Web Image Dataset for Event Recognition FACE (WIDER FACE) dataset. ALM detects faces in COCO 2017 dataset and provides bounding boxes automatically with very high accuracy. The generated new labels by ALM are merged to the existing labels in COCO 2017 dataset automatically. 4,110 hours human effort is saved using this system. The expanded COCO 2017 dataset is evaluated quantitatively and visually to ensure the efficiency of the proposed auto-labeling system.