Compute-resource Allocation for Motion Estimation in Real-time Video Compression

We present a novel method for allocating computational resources in motion estimation for real time video compression, in order to attempt to minimize prediction error over the whole image. The method makes decisions about how much compute power is allocated to motion estimation for each particular block based upon past observations of the input and output data of each task. We present preliminary results that demonstrate the ability of our method to improve the resulting quality of encoded video under a realtime computational constraint.