Yaw approximated by X-translation


We take an input image and blur it using a yaw camera motion. We solve for the Motion Density Function (MDF) restricted to x-translation with the latent image fixed to the known ground truth image.  Finally, we deblur the image using the kernels generated by the MDF. We measure the PSNR of the deblurred image w.r.t. the ground truth. A higher PSNR indicates that x-translation performed a good approximation of the yaw motion.

Following is a plot of the PSNR as we vary the focal length. We decreased the yaw motion with the increasing focal length to keep the kernel size similar across all tests. This is to avoid confounding the PSNR metric by the non-blind deconvolution errors resulting from different kernel sizes.



The PSNR improvement levels out quickly and the translation-only kernels start to approximate the yaw kernels quite well. Hence this approximation holds well for not-very short focal lengths.

Notice the image corners where translation-only model has problems.

Focal-length Blurred Image Deblurred Image Actual blur kernels Recovered blur kernels using ground truth image as initialization
20mm


30mm


40mm


50mm


70mm


100mm


120mm


150mm
200mm
300mm
400mm