Research Of Object Proposals Based On RGB-D Data

Object proposals is a generic objectness measure, quantifying how likely it is for an image window to contain an object of any class. It can be used for reducing searched object windows and improving accuracy in the object detection and recognition applications. In recent years, in the field of object recognition, the paradigm relies on perceptual grouping has dominated the object recognition challenge. Since the power of this paradigm is critically dependent on the accuracy and the number of object proposals, an increasing body of research has delved into the problem. Meanwhile, the availability of inexpensive RGB-D sensors, such as Microsoft Kinect, Apple PrimeSense and Intel RealSense has greatly simplified some common challenges in vision. The increase of the depth cue will also promote the development of object proposals, and can further reduce the search windows for object recognition. However, the existing algorithms are not enough for using of the depth information. In this paper, we mainly study the effective use of depth information in object proposals. But most of the data commonly afforded in image processing applications is still RGB version. And for the single image, we propose a method to estimate the depth based on structured learning, In detail, the main points of this work would be as follows. (1) We propose an approach of object proposal via RGBD images, a) We design edge related cues and depth objectness cues by depth information, b) Improving edge based object proposal using our new edge cues, c) Proposing a depth based object proposal algorithm, d) Using a Bayesian framework to combine them. We demonstrate that the combined objectness measure performs better than any cue alone on the NYU Depth dataset, and also outperforms traditional objectness based on RGB image. (2) For a single image, we propose a method to estimate the depth based on structured learning, a) We propose a multi-scale structured learning framework based on random forest, b) Proposing a depth labels discrete method to solve computational problem of information gain, c) Proposing an out-of-bag based feature optimization method via structured forest. We achieve a better estimation effect on the Make3d and NYU Depth v2 dataset, and our algorithm is faster than state-of-the-art approaches.