Photometric stereo offers the possibility of object shape reconstruction via reasoning about the amount of light reflected from oriented surfaces. However, in murky media such as sea water light interacts not only with the scene but also with the medium particles. The resulting image formation model is complex and hard to optimize directly for scene orientation and albedo. The effectiveness of previous solutions has been limited since they were based on unrealistic assumptions about the imaging conditions or on external hardware for estimating some of the unknown model variables.

In this thesis we propose a complete Photometric Stereo framework that yields useful reconstructions within a wide range of conditions in murky water. First, we show how the problem can be simplified by taking account of the additive light component that is backscatterd by the particles. A backscatter compensation method is proposed, based on the observation that the respective light signal varies smoothly across the sensor according to the pixel position with respect to the light source but is scene-depth independent. This makes its approximation experimentally appropriate using only the captured images and results in shape reconstruction similar to that in clean water. Second, we propose an uncalibrated method for solving Photometric Stereo in the presence of near-field illumination within murky water. In this case, the incident illumination on the scene is non-uniform as it varies according to the scene depth and the attenuation coefficient of the medium. We propose an algorithm for optimizing the resulting complex system of equations without resorting to external equipment or calibration. Finally, we design a novel Photometric Stereo system with dynamic lighting that can predict the validity of different photometric models in uncontrolled imaging conditions. Both our and previous approaches have been based on simplifications of the original-complex image formation in murky water. The effectiveness of such simplified models differs according to the imaging conditions and it can not be easily estimated as both the scene and the environment are usually unknown. Our proposed system can predict the validity of such photometric models, based on the observation that the change in estimated surface normals as the light sources position is varied with respect to the camera is a direct estimate of a model's effectiveness. In this way, the system can approximate the true reconstruction error and adapt automatically some critical factors in order to maximize the reconstruction quality, such as the camera-scene distance, the light sources position and the photometric model. Our work is evaluated through extended numerical simulations and real experiments in a water tank and real port water, and yields detailed reconstructions over a wide range of distances and underwater visibilities.