3D Depth Recovery Of Natural Scenes And Applications

With the rapid development of display technology and the growth of our daily life needs,3D stereoscopic display technology 3D stereoscopic display technology has already set off a new technological revolution in the field of graphics and images displaying,and becomes the latest and most cutting-edge high technology in imaging industry.It attracts people's attention by its new,unique and singular expression manner,true and strong visual impact,and comfortable,beautiful environment.Meanwhile,3D stereoscopic display technology is widely used in many areas.However,most of existing images and videos are still in 2D form.So how to recovery 3D depth information from the2 D images or videos becomes an important task in the field of stereoscopic 3D display.With the depth information,2D scenes can be easily converted into stereoscopic form.In this paper,we focus on 3D depth recovery and applications of 2D scenes.The major contributions of the paper are:1.We introduce an effective method for depth estimation from a single natural image using defocus cues,which is inspired by the observation that defocusing can significantly affect the spectrum amplitude at the object edge locations in an image.By considering that there may exist noises in the initial recovered depth map,we propose an efficient edge preserving image smoothing method,which can smooth the detailed texture of an image and preserve the important geometric edges.By using this smoothing method,we can get a smoother depth map.2.We first present a RGB image saliency detection method which is based on the image background prior and color spatial distribution.By adding depth information,we then propose a depth aided RGBD image saliency detection method.Experimental results show that our method can get higher precision and recall than previous methods on different datasets.3.By using the depth information,we propose a online action recognition framework.In our framework,we use 3D points of the human body extracted from depth map to formulate the action descriptors.Moreover,our method recognise human action in a continuous manner by online updating the action descriptors.Compared with previous methods,our results show higher accuracy,low latency.