Data Driven Coded Aperture Design for Depth Recovery

摘要: Inserting a patterned occluder at the aperture of camera lens has been shown to improve recovery depth map and all-focus image compared fully open aperture. However, design pattern plays very critical role. Previous approaches for designing codes make simple assumptions on distributions obtain metrics evaluating codes. real images may not follow those hence designed code be optimal them. To address this drawback we propose data driven approach learning recover from single coded image. We two stage architecture where, in first simulate training dataset maps second using deep neural network. demonstrate that our learned performs better than previously even proposed by previous approaches.