Cut and Paste: Generate Artificial Labels for Object Detection

Abstract

In the domain of object detection, region proposal, feature extraction, recognition and the localization are the main three tasks. The end-to-end detection models by integrating the three parts together to simplify the structure of network and accelerate the process of training and detection. While the issues of illumination change, object deformation and scale change undermine the performance of detection methods largely. To promote the object detection accuracy rate and boost the detection speed simultaneously, we propose a new method of data augmentation. Different from the traditional methods, our method can increase the training data largely and be free from overfitting to some extent. With the new method, the abstraction ability of models improves a lot, the model has better performance to multiscale objects detection, and also has a stronger distinguishing ability in complex background.