GPU-accelerated high-resolution image stitching with better initial guess

High-resolution image processing is crucial for emerging applications in virtual reality. The trade-off between robustness/accuracy and speed pose a challenge to high-resolution image stitching. This paper accelerates the state-of-the-art image stitching algorithm by using the CUDA toolkit. We also use the mesh coordinates as an initial guess for solving warping function for an additional increase in speed. Experimental results show that our implementation achieves comparable quality to the state-of-the-art work, while our implementation is almost 2x faster on high-resolution images (3264×2448). © 2018 IEEE.