Unsupervised Object Proposal Using Depth Boundary Density and Density Uniformity

Object proposal that detects candidate bounding boxes of objects in images is an effective way of accelerating object recognition in the robot/computer vision area. We propose an accurate and fast object proposal method using depth images. Existing proposal methods can be roughly divided into two categories: window scoring and object region extraction. The window scoring methods usually have higher efficiency than object region extraction methods. The previous methods using RGB images detect an excessive number of boxes due to edges of texture objects. These methods also may misdetect overlapping objects as one candidate bounding box. To tackle these problems, we propose a novel and effective objectness measure using depth images. The proposed method evaluates objectness by using depth boundary density difference between inner and outer regions of a candidate bounding box. We also consider the uniformity of the outer boundary density in a candidate bounding box to divide overlapping objects into individual candidate bounding boxes. Our reasonable assumption here is that the depth boundary of an object has a closed loop. Our experiments show significant performance gains over existing RGB and RGB-D object proposal methods on the challenging toy-dataset [1] of complex crowded scenes.