We propose a rendering method for efficiently computing the transmitted caustics produced by a glass panel with arbitrary surface deformations, characteristic of old glass used in 3D reconstructions in virtual heritage. Using Fermat's principle of least time, we generalize the concept of Next Event Estimation to allow light sampling through two displaced refractive interfaces, which amount to numerically finding all stationary points of an objective function. Our work allows for an efficient estimation of the caustic while staying inside a standard Monte Carlo pathtracing framework. Our specific geometrical context allows our solver to converge significantly faster than the more general method Specular Manifold Sampling, while scaling well with the number of panels present in the scene.