In this paper we present a methodology for hand pose estimation from a single image, combining bottom-up and top-down processes. A fast bottom-up algorithm generates, from coarse visual cues, hypotheses about the possible locations and postures of hands in the images. The best ranked hypotheses are then analysed by a precise, but slower, top-down process. The complementary nature of bottom-up and top-down processes in terms of computational speed and precision permits the design of pose estimation algorithms with desirable characteristics, taking into account constraints in the available computational resources. We analyse the trade-off between precision and speed in a series of simulations and qualitatively illustrate the performance of the method with real imagery.