High-Dimensional Camera Shake Removal With Given Depth Map

Camera motion blur is drastically nonuniform for large depth-range scenes, and the nonuniformity caused by camera translation is depth dependent but not the case for camera rotations. To restore the blurry images of large-depth-range scenes deteriorated by arbitrary camera motion, we build an image blur model considering 6-degrees of freedom (DoF) of camera motion with a given scene depth map. To make this 6D depth-aware model tractable, we propose a novel parametrization strategy to reduce the number of variables and an effective method to estimate high-dimensional camera motion as well. The number of variables is reduced by temporal sampling motion function, which describes the 6-DoF camera motion by sampling the camera trajectory uniformly in time domain. To effectively estimate the high-dimensional camera motion parameters, we construct the probabilistic motion density function (PMDF) to describe the probability distribution of camera poses during exposure, and apply it as a unified constraint to guide the convergence of the iterative deblurring algorithm. Specifically, PMDF is computed through a back projection from 2D local blur kernels to 6D camera motion parameter space and robust voting. We conduct a series of experiments on both synthetic and real captured data, and validate that our method achieves better performance than existing uniform methods and nonuniform methods on large-depth-range scenes.