Why merge a burst instead of one long low-light exposure?

WHAT THIS TESTS The interviewer wants the signal-processing rationale for bursts and a robust merge that survives camera shake and moving subjects.

A GOOD ANSWER COVERS Why burst beats one long exposure: a single long exposure blurs any motion, may clip highlights, and can saturate; many short exposures freeze motion and preserve highlights, and because read noise and shot noise are largely independent across frames, averaging aligned frames raises the signal-to-noise ratio by about the square root of the frame count. The algorithm picks a reference frame, then aligns every other frame to it. Global alignment alone cannot handle parallax and local motion, so alignment is usually hierarchical and tile-based, estimating a local shift or flow per patch. During merge you accumulate aligned pixels, but to avoid ghosting you weight each contribution by how well it matches the reference; where a moving object or a misalignment causes disagreement, you reduce its weight or fall back to the reference, trading a little noise reduction for artifact-free output.

COMMON WRONG ANSWERS Plain averaging with no alignment, producing blur. Ignoring local object motion and getting ghosts. Claiming SNR scales linearly with frame count rather than with its square root.

LIKELY FOLLOW-UPS Frequency-domain merging as in Google's HDR+, choosing the sharpest reference frame, handling alignment failure, and combining burst merge with tone mapping.

ONE CONCRETE EXAMPLE Google's HDR+ captures a burst of underexposed raw frames, aligns them in tiles to a chosen reference, and merges in a way that suppresses noise while rejecting tiles where a person moved, then brightens the result, yielding a clean low-light photo without the blur of a one-second exposure.