Burst photography for high dynamic range and low-light imagingn mobile cameras6pt] Supplemental Material

At the finest scale of our coarse-to-fine alignment strategy, we require an alignment technique which performs well given large tile sizes but a very small search radius. In this context, techniques such as our previously-described fast subpixel L2 alignment, or even simpler techniques such as phase correlation [Kuglin and Hines 1975] are outperformed by a well-implemented brute-force procedure which minimizes absolute residuals. We use absolute residuals instead of squared residuals because they map well to low level computer architectures. In contrast with squared residuals, absolute residuals require fewer bits for the same input data, enabling higher throughput. On ARM architectures, there is in fact an operation which computes an absolute difference and then accumulates, making brute-force computation of L1 distances very computationally light. This approach scales quadratically with the search radius, making it most appealing in contexts where the search radius is small.