SINGLE-SHOT OBJECT DETECTORS AND THE FUTURE DIRECTION OF OBJECT DETECTION

Abstract

Object Detection is a core task in computer vision and essential for many downstream tasks/ap- plications such as robotics, self-driving cars, satellite image analysis, and e-commerce. In this dissertation, I present a series of works on object detection, from efficient single-shot object detection to improving image segmentation. In the first part, I propose a series of works on Single-Shot Detection(SSD); SSD is widely cited for speed performance and great potential for real-time applications. With this first work, Single-Shot Detection, I demonstrates that classification and regression of object detection can be performed directly without feature resampling. The proposed approach detects objects in an im- age in a single convolution neural network. Additionally, the network combines prediction from multiple feature maps with different resolutions to naturally handle objects of various sizes. In the second work, Deconvolution Single-Shot Detection, I add a top-down connection introducing additional contextual information to the current object detection framework and improving accu- racy, especially for small objects. I developed the last work, Learning Mask To Improve Object Detection, to increase the performance of the state-of-the-art single-shot detector, RetinaNet, in three ways: integrating the instance mask prediction(the first single-shot detector to do so), creat- ing an adaptive and more stable loss function, and including additional hard examples in training. I named the resulting method RetinaMask; the detection component of RetinaMask has the same computational cost as the original RetinaNet but is more accurate. In the second part, I present a new operator called the Instance Mask Projection (IMP), which projects the predicted instance segmentation as part of feature maps for semantic segmentation. Our experiments show the effectiveness of IMP on both clothing parsing (with complex layering, large deformations, and non-convex objects), and on street scene segmentation (with many overlapping instances and small objects).