Treffer: Proposal-free framework for object instance segmentation using deep learning

M.Phil. ; Object detection, which deals with finding instances of semantic objects of predefined classes (e.g., humans, cars, or dogs) in digital images and videos, is one of the fundamental tasks in the field of computer vision. In the research community, object detection is the base for many other tasks, such as 3D detection, human identification, relationship,and etc, Also, it is the key element and has been deployed in real-world applications including smartphone apps, autonomous driving cars, and intelligent surveillance cameras. ; The task definition of object detection is simple. it is nothing but just to locate objects and tell the categories. However, it is quite challenging. Due to large variations in viewpoints, poses, occlusions, lighting conditions and background, object detection is challenging. The visual cues from multiple support regions of different sizes and resolutions are complementary in classifying a candidate box in object detection. In natural digital images, there exists a complex structure between objects, e.g, it is very likely that the water bottles should be on tables. But how to model this relationship is still an open question. Recent years, several deep learning based detection frame works are proposed. Thanks to the powerful representations, we achieve a significant improvement. However, most of them are quite time-consuming. How to make existing detectors more efficient, or even propose a new one is vital. ; In this dissertation. we address these challenges from four aspects to make object detection scalable to real-world data and applications: ; •First, how to learn better feature representations and classifiers in object detection. Feature matters. We propose to take features of different resolutions and support regions and pass messages to each other to validate their existence through the bi-directional structure. And the gate function is used for controlling the message passing rate among these features. ; •Second, we introduce graphic model. Graphic model is good at ...