AI Glossary / Computer vision and multimodal AI / Intermediate

Burst photography

Definition

Burst photography captures a sequence of images in rapid succession. In computational photography, the frames can be aligned, selected or merged to improve a final photograph, for example by reducing noise, preserving highlights or exploiting information distributed across multiple captures.

What it means

A burst can simply give a photographer several moments to choose from. Computational burst imaging goes further by combining information. Several short exposures may offer advantages over one long exposure when camera or subject motion would otherwise blur the image.
The frames are not automatically identical. Camera shake, moving subjects, exposure differences and noise can change their content. A useful merge must distinguish corresponding scene information from changes that should not be averaged together.
Burst capture does not guarantee a better result. It consumes time, memory and computation, and the benefit depends on scene motion, exposure strategy, sensor behavior and the merging method.

How it works

Capture a sequence using a chosen exposure and timing strategy. Select a reference frame and estimate how other frames align with it. Identify moving or inconsistent regions, then combine reliable measurements using suitable weights or robust rules. Apply the remaining image-processing steps needed for the intended output.
For independent zero-mean noise and perfectly aligned measurements of an unchanged scene, averaging N frames reduces noise variance by a factor of N. Real systems deviate from those assumptions: noise can be signal-dependent, fixed-pattern errors can persist and motion can limit usable frames. Equal-exposure bursts and exposure-bracketed sequences also solve somewhat different capture problems.

Three practical examples

• A low-light imaging pipeline aligns several short exposures to reduce noise without relying entirely on one long exposure.
• A photographer selects a sharp frame from a rapid sequence of a moving subject.
• A computational camera merges suitably exposed frames to retain useful shadow information while avoiding clipped highlights.

Think of it this way

Think of collecting several imperfect observations of the same scene and combining the parts that agree. Random errors can cancel, while repeated structure reinforces. The analogy works only when the observations can be aligned and the scene is sufficiently consistent. Averaging a moving hand across frames can create multiple outlines rather than a clearer hand.

Worked example

Suppose four perfectly aligned frames measure the same static pixel with independent noise having standard deviation 8 units in each frame. Averaging the four values gives noise variance 8²/4=16 and standard deviation 4. The signal level remains unchanged, so its signal-to-noise ratio doubles under these assumptions.
Now suppose one frame contains a moving object at that location. Its value is not merely another noisy measurement of the same signal. Blindly averaging it can create a ghost or dilute the object boundary.
A robust burst method may reject that contribution or use a reference-frame value in the inconsistent region. The calculation therefore explains a potential noise benefit, not a guaranteed improvement for every pixel or a promise that four real frames always double photographic quality.

Illustrative case study

Fictional case: Rowan Cameras evaluates a burst merge on indoor scenes. Static furniture looks cleaner, but a waving person develops faint duplicate edges. The first quality report focuses on noise measurements from still regions.
The team adds moving subjects, alignment failures and highlight clipping to its evaluation. It checks which frames contribute to each region and whether fallback behavior preserves a usable reference image. It also measures capture and processing latency. This case illustrates the trade-offs of multi-frame imaging rather than a claimed result for any commercial camera or a guarantee that burst capture eliminates motion artifacts.

Common misconceptions and mistakes

A burst is not necessarily exposure bracketing; many pipelines use repeated similar exposures. More frames do not always improve the result if alignment, motion or latency becomes problematic.
Noise averaging does not remove every fixed-pattern error or recover clipped information from frames that all saturated at the same location. Sharpness, noise, dynamic range and resolution are different properties. A cleaner image is not necessarily a more faithful one if merging invents boundaries or erases transient content. Single-frame and burst comparisons should account for total capture conditions.

Business use

Burst imaging can support product photography, mobile capture and specialized inspection when a sequence provides more useful information than one frame. It can improve the chance of obtaining a usable image, but its cost and failure modes should match the workflow.
Evaluate static and moving scenes, low light, highlights and repeated textures. Record exposure strategy, frame count, reference selection and merge behavior. Measure end-to-end latency and storage requirements as well as image quality. For evidence or measurement tasks, retain original frames and processing metadata where appropriate so users can distinguish captured content from the final fused representation.

People also ask

Does burst always mean merging?

No. It can also mean selecting one frame from a rapid sequence.

Does averaging always remove motion blur?

No. Motion and misalignment can introduce blur or ghosting.

Are all burst frames differently exposed?

No. Equal-exposure and bracketed approaches both exist, with different goals.

Check your understanding

Four frames show the same background, but a person changes position between them. Should every pixel be averaged without checking consistency?

Answer: No. Moving regions are not repeated measurements of an unchanged signal. The merge needs motion-aware handling or a suitable fallback to avoid ghosting and detail loss.

Related terms

• Exposure bracketing
• Burst super-resolution
• Image alignment
• HDR imaging
• temporal denoising
• ghosting
• reference frame
• signal-to-noise ratio

Learn next

Learn image alignment and noise averaging next. Compare a static burst with one containing motion, then inspect where averaging helps and where it fails. Distinguish noise reduction from increased dynamic range or resolution when evaluating a multi-frame pipeline.

• Noise reduction
• Dynamic range

More in Computer vision and multimodal AI

• Box filter - A box filter is an image filter that combines values within a rectangular neighborhood using equal weights.
• Brown - Conrady model - The Brown - Conrady model is a camera-distortion model that describes departures from ideal perspective projection using radial and tangential terms.
• Bundle adjustment - Bundle adjustment is an optimization procedure that jointly refines camera parameters and, in many reconstruction settings, three-dimensional point...
• Calculus of variations - The calculus of variations is a branch of mathematics for finding functions that make a functional stationary or optimal.
• Calibration frames (bias, dark, flat) - Calibration frames are reference images used to estimate and correct systematic sensor or optical effects.
• Camera path - A camera path is the trajectory of a camera's position and orientation over time, or an image-based representation of that motion.

Sources

Samuel W. Hasinoff and colleagues, Burst photography for high dynamic range and low-light imaging on mobile cameras: https://people.csail.mit.edu/hasinoff/pubs/HasinoffEtAl16-hdrplus.pdf Authors' dataset: https://hdrplusdata.org/ Primary paper and dataset excerpts consulted. The noise calculation is original and assumption-bound; full imaging and editorial review remain pending.

Previous: Burst buffer Next: Burst super-resolution