Abstract
We describe a method for recovering the irradiance underlying a collection ofimages corrupted by atmospheric turbulence. Since supervised data is oftentechnically impossible to obtain, assumptions and biases have to be imposed tosolve this inverse problem, and we choose to model them explicitly. Rather thaninitializing a latent irradiance ("template") by heuristics to estimatedeformation, we select one of the images as a reference, and model thedeformation in this image by the aggregation of the optical flow from it toother images, exploiting a prior imposed by Central Limit Theorem. Then with anovel flow inversion module, the model registers each image TO the template butWITHOUT the template, avoiding artifacts related to poor templateinitialization. To illustrate the robustness of the method, we simply (i)select the first frame as the reference and (ii) use the simplest optical flowto estimate the warpings, yet the improvement in registration is decisive inthe final reconstruction, as we achieve state-of-the-art performance despiteits simplicity. The method establishes a strong baseline that can be furtherimproved by integrating it seamlessly into more sophisticated pipelines, orwith domain-specific methods if so desired.