Building a 3D model of the operating field is the first stage to access the registration of pre-operative surface models in per-operative context. We propose a method to learn the 3D model of the organs observed by a stereoscopic endoscope. This method is general and requires no a priori knowledge of the position of the instruments, whether robotized or manually-controlled, present in the field of view. The result of the learning process allows the removal of the instruments from the surgeon?s field of view: "diminished reality". This work prefigures the 3D integration by augmented reality of preoperative organs model (e.g. arteries) in stereoscopic images. Experiments are currently being performed to validate our approach using endoscopic sequences from the da Vinci^TM surgical system.