Fast GPU Approximation of EPZS Motion Estimation Using Branching

This paper studies the adaptation of one of the most popular motion estimation methods to a massively parallel environment. To achieve the performance benefits that this and other related methods exhibit in CPU, a proper use of the GPU branching capabilities is required. We prove this claim and the viability of the branching approach, and achieve performance benefits of over x7 compared to other fast GPU approaches. Our implemented method works very close in encoding efficiency to the reference CPU encoder for small frame sizes and high definition sequences with high bitrates, but suffers at low bitrates in high definition sequences, a fact that is common to other fast GPU implementations, but is stronger in this case due to the nature of the EPZS method. The implemented approach achieves a motion estimation performance of 722 fps on 1080p sequences using a NVIDIA GTX 580 GPU.