A preliminary study on extracting objects in sketches

Humans have the incredible ability to interpret complex sketch images, but it remains a challenging task for computers to do the same thing. Researchers have made great progress on nature image detection and recognition, but little research has been done on object detection in sketch images. In this paper, we demonstrate that object extraction can be possibly conducted using a Single-Shot Multibox Detector(SSD) frame-work, without the guidance of segmentation information or user interaction. We train and test our model with a synthetic dataset based on TU-Berlin sketch dataset. Experiments on the synthetic dataset show reasonable object detection and recognition results in sketch images.