High Quality Novel View Synthesis Based on Low Resolution Depth Image and High Resolution Color Image Jui-Chiu Chiang, Zheng-Feng Liu, and Wen-Nung Lie National Chung Cheng University Taiwan 2012/11/11 1 Outline 2012/11/11 Introduction Our hybrid camera system The proposed depth processing algorithm Experimental results Conclusion 2 Color + Depth in 3DTV/FTV Depth image, in combination with DIBR (Depth Image-Based Rendering) technique, is important in multi-view 3DTV/FTV applications. Color image Depth image DIBR view9 2012/11/11 view8 view7 view6 view5 view4 view3 view2 view1 3 Depth acquisition From stereo matching via disparity estimation From active photo-electrical sensing Kinect 2012/11/11 Unreliable for textureless regions Computationally intensive Lower resolution Higher acquisition speed and accuracy SR4000 ToF 4 Hybrid camera systems High-resolution color cameras + Low-resolution depth camera (mis-alignment between their optical axis) [12]: Use warped data of depth camera to refine ROI disparity of the dual-eye camera Proposed: one color image is used to refine the warped (to both sides) and up-sampled depth images SR4000 Camera L 2012/11/11 Camera R 5 Processing diagram of our system left color image Left preliminary Left high-quality high-resolution depth high-resolution depth image processing image processing Right preliminary Right high-quality high-resolution depth image processing high-resolution depth Left depth image central depth image right color image 2012/11/11 Right depth image image processing 6 Generation of preliminary highresolution depth image Purpose: to generate high-resolution depth image at the left and right sides Depth image from depth camera Pre-processing depths to remove noise Calibrated camera parameters Warping and up-sampling depth image to the reference views Preliminary high resolution depth images Post-processing warped depths to remove incorrect depth values 2012/11/11 7 Pre-processing of the depth map Low resolution to high resolution depth warping Forward warping with a block size of 55 pixels Post-processing by morphological operation 2012/11/11 33 median filter Morphological opening 8 Generation of high quality highresolution depth image Color-image-guided depth map refinement Segment n and Db(x,y) 1 Sn depth value is correct? Yes depth value is unchanged 0 No depth value D(x,y) with Db(x,y)=0 inside the segment n is averaged and used as the updated depth value depth value is correct? No Yes depth value D(x,y) with Db(x,y)=1 inside the segment n is averaged and used as the updated depth value depth value is unchanged 2012/11/11 9 Binarizing the warped depth image to have foreground/background classification Color segmentation by mean-shift algorithm 2012/11/11 Label each segment as FG or BG Refine the corresponding depth values according to the segment categorization Bilateral filtering is performed for the edge parts of the modified depth image 10 Experiments – system configurations 2012/11/11 Color image: 640x480 Depth image: 176x144 (SR4000) Intel Core 2 Duo Q6600 2.33 GHz, and DDR2 800 4GB 11 Preliminary high-resolution depth image After depth warping 2012/11/11 (a) (c) (e) (b) (d) (f) Traditional depth warping After morphological operation 12 High-quality high-resolution depth image After correction Meanshift (a) (c) (e) Incorrect pixels (b) 2012/11/11 (d) (f) Bilateral filtering 13 A series of generated depth maps 2012/11/11 14 Virtual view synthesis based on various depth maps (a) (c) (e) (b) (d) (f) Hole filled 2012/11/11 Bilateral filtering Bilateral filtering on edges Bilateral filtering on 15 edges + Gaussian Speed performance Pre-processing Post-processing Procedure Time (sec) Depth Warping 0.117 Color segmentation 0.742 Bilateral filtering on edges 0.352 Gaussian filtering 0.104 Novel view synthesis by DIBR (including hole filling) 2012/11/11 0.144 16 Conclusion A hybrid camera system consisting of highresolution color camera and low-resolution depth camera is proposed A high-quality high-resolution depth images can be obtained for the left and right eyes – multiview video plus depth (N=2 channels). Future work 2012/11/11 Combine Kinect device and stereo camera to get better depth image and cheaper cost 17