Choose an algorithm

Choose the simplest model that describes the variation in your data, then test what happens when its assumptions fail.

You want to…	Start with	Check first
Reduce ordinary image noise	Gaussian	Can you afford to blur boundaries?
Remove isolated specks	Median	Must thin lines survive?
Preserve strong boundaries while smoothing	Bilateral or guided filtering	Does the guide contain the right structure?
Separate a bright object	Threshold	Is illumination uniform?
Segment under uneven lighting	Adaptive threshold	What neighbourhood captures illumination without following texture?
Count separate binary regions	Connected components	Are touching objects already separated?
Measure object outlines	Contours	Which boundaries and holes should count?
Find an unchanged patch	Template matching	Are scale and orientation stable?
Match a textured planar object	ORB or SIFT, then homography	Is one plane a valid model?
Follow points between nearby frames	Lucas-Kanade	Are the patches visible and textured?
Estimate dense motion	Farneback or DIS	How will you identify uncertain or occluded pixels?
Refine an existing alignment	ECC	Is the initial warp close enough?
Estimate translation of the whole frame	Phase correlation	Does most of the content share one translation?
Read a marker or code	ArUco or QR	Are the code’s cells resolved clearly?
Estimate known-object pose	PnP	Are camera calibration and point correspondences correct?
Run a learned detector	DNN	Does the model’s operator set work in this build?

Evaluate the assumption, not just the output

A crisp mask may still identify the wrong object. A low alignment error may come from a repeated texture. A smooth depth map may contain invalid correspondences. Keep confidence, coverage and model fit separate from the visual attractiveness of the result.

Use a positive control that should succeed, a nearby difficult case and a case that should be refused. For image reconstruction, exclude genuinely unknown pixels from numerical error while reporting how many are unknown.