The computation of high quality animation sequences is expensive. Generation time for each frame can take a few hours Recently, Image-Based Rendering methods have been proposed to solve this problem. As these techniques obtain new arbitrary views from precomputed ones at low cost, walkthroughs may be computed faster. Consequently, the selection of the precomputed images is a very important step. The initial set of keyframes should fulfill two requirements, it must be small but provide as much information as possible on the scene. In this paper we review several keyframe selection strategies and then we propose a new method based on entropy that achieve similar, and in some cases better, results.