Depth map recovery from videos

depth map recovery;computer vision;motion segmentation

Wang, Jiqin ; Forsyth ; David A.

关键词: depth map recovery;    computer vision;    motion segmentation;   

Others  :  https://www.ideals.illinois.edu/bitstream/handle/2142/50664/Jiqin_Wang.pdf?sequence=1&isAllowed=y

美国|英语

来源: The Illinois Digital Environment for Access to Learning and Scholarship

【 摘 要 】

The depth map of a video is a very important piece of information. Recovering the depth map of a video expands a 2D video into its 3rd dimension, and creates new possibilities, such as, object insertion, conversion to 3D, shallow depth of field simulation.In this work, we introduce our approach of recovering depth maps from a video sequence with a moving camera and moving objects. Our approach isolates moving objects of each frame and estimates the depth of the scene and the moving objects separately. It takes advantage of the fact that the surfaces that belong to the same object share similar optical flow angles, and have smooth optical flow angle gradients, that can be exploited to recover object boundaries, thereby isolating moving objects from the static part of the scene.\\It recovers the relative depth of the static part of the scene by calculating the likelihood of a pixel belonging to the farthest background using the magnitude of the optical flow and recovered 3D points. It then estimates the depth of moving objects by finding a statistically most likely actual size of the object and converting the actual size to its actual depth. Finally, we reinsert the estimated depth moving object into the estimated depth of the rest of the scene.