Learning an Object Detector Using Zoomed Object Regions

摘要

The single shot multi-box detector (SSD) is one of the first real-time detectors, which uses a convolutional neural network (CNN) and achieves the state-of-the-art detection performance. However, owing to the semantic gap between each feature layer of CNN, the SSD has a room for improvement. In this paper, we propose a novel training scheme to enhance the performance of the SSD. In object detection, ground truth (GT) box is a bounding box enclosing an object boundary. To improve the semantic level of the feature map, we generate additional GT boxes by zooming in to and out from the original GT boxes. Experimental results show that the SSD trained with our scheme outperforms the original one on public dataset.