Images of underwater scenes suffer from poor contrast. Water-induced contrast decay varies across the scene and is exponential in the depths of scene points, which prevents standard computer vision algorithms from operating properly. In this paper we show how to overcome this problem by adapting an existing model of light propagation in the (foggy) atmosphere to describe the behavior of light in liquid media. By integrating the resulting model within a dense stereo algorithm, we recover disparity maps of scenes immersed in water from pairs of images of these scenes acquired from distinct viewpoints. Experiments performed with real underwater images of various degrees of turbidity show that the use of a physically-based light-propagation model allows one to reconstruct underwater scenes more accurately than with standard stereo algorithms alone.