This paper analyzes and evaluates recent blind deconvolution algorithms. Blind deconvolution aims to recover a sharp image from a blurred one when the blur kernel is unknown. The paper highlights the limitations of the maximum a posteriori (MAP) approach, which often favors the no-blur explanation. It shows that while the MAP approach for both image and kernel (MAP_{x,k}) fails, a MAP approach for the kernel alone (MAP_{k}) can accurately recover the true blur kernel. This is due to the strong asymmetry between the dimensionalities of the image and the kernel, where the kernel's dimensionality remains small even as the image size increases. The paper evaluates various blind deconvolution algorithms using ground-truth data. It demonstrates that the variational Bayes approach outperforms other methods. The study also shows that the shift-invariant blur assumption used by most algorithms is often violated in real-world scenarios, as realistic camera shake includes in-plane rotations. The paper discusses the importance of choosing an appropriate estimator over the prior. While sparse priors are useful, the key to successful blind deconvolution lies in the thoughtful choice of estimator. The paper also shows that even with a weak Gaussian prior, blind deconvolution can be performed effectively. The paper concludes that future research should focus on improving existing estimators rather than relying solely on advanced priors. It also emphasizes the need for blur models that relax the spatially uniform blur assumption, which is often unrealistic in practice. The study provides insights into the challenges and potential improvements in blind deconvolution algorithms.This paper analyzes and evaluates recent blind deconvolution algorithms. Blind deconvolution aims to recover a sharp image from a blurred one when the blur kernel is unknown. The paper highlights the limitations of the maximum a posteriori (MAP) approach, which often favors the no-blur explanation. It shows that while the MAP approach for both image and kernel (MAP_{x,k}) fails, a MAP approach for the kernel alone (MAP_{k}) can accurately recover the true blur kernel. This is due to the strong asymmetry between the dimensionalities of the image and the kernel, where the kernel's dimensionality remains small even as the image size increases. The paper evaluates various blind deconvolution algorithms using ground-truth data. It demonstrates that the variational Bayes approach outperforms other methods. The study also shows that the shift-invariant blur assumption used by most algorithms is often violated in real-world scenarios, as realistic camera shake includes in-plane rotations. The paper discusses the importance of choosing an appropriate estimator over the prior. While sparse priors are useful, the key to successful blind deconvolution lies in the thoughtful choice of estimator. The paper also shows that even with a weak Gaussian prior, blind deconvolution can be performed effectively. The paper concludes that future research should focus on improving existing estimators rather than relying solely on advanced priors. It also emphasizes the need for blur models that relax the spatially uniform blur assumption, which is often unrealistic in practice. The study provides insights into the challenges and potential improvements in blind deconvolution algorithms.