Solving Multi-view Stereo and Image Restoration using a Unified Framework

Abstract

Estimating camera pose and scene structures from seriously degraded images is challenging problem. Most existing multi-view stereo algorithms assume high-quality input images and therefore have unreliable results for blurred, noisy, or low-resolution images. Experimental results show that the approach of using off-the-shelf image reconstruction algorithms as independent preprocessing is generally ineffective or even sometimes counterproductive. This is because naive frame-wise image reconstruction methods fundamentally ignore the consistency between images, although they seem to produce visually plausible results. In this thesis, from the fact that image reconstruction and multi-view stereo problems are interrelated, we present a unified framework to solve these problems jointly. The validity of this approach is empirically verified for four different problems, dense depth map reconstruction, camera pose estimation, super-resolution, and deblurring from images obtained by a single moving camera. By reflecting the physical imaging process, we cast our objective into a cost minimization problem, and solve the solution using alternate optimization techniques. Experiments show that the proposed method can restore high-quality depth maps from seriously degraded images for both synthetic and real video, as opposed to the failure of simple multi-view stereo methods. Our algorithm also produces superior super-resolution and deblurring results compared to simple preprocessing with conventional super-resolution and deblurring techniques. Moreover, we show that the proposed framework can be generalized to handle more common scenarios. First, it can solve image reconstruction and multi-view stereo problems for multi-view single-shot images captured by a light field camera. By using information of calibrated multi-view images, it recovers the motions of individual objects in the input image as well as the unknown camera motion during the shutter time. The contribution of this thesis is proposing a new perspective on the solution of the existing computer vision problems from an integrated viewpoint. We show that by solving interrelated problems jointly, we can obtain physically more plausible solution and better performance, especially when input images are challenging. The proposed optimization algorithm also makes our algorithm more practical in terms of computational complexity.