AI blur and blink detection: what fools each check

Sharpness and eyes open are the two checks photographers argue with most, because both measure something real and neither measures the thing you asked. Here is how each number is actually produced, a catalogue of the frames that fool each one, and the handful of habits that keep you out of trouble.

Why these two cause the most arguments

An AI culling tool runs four or five measurements over your files. We took the whole set apart in how AI picks your best photos. This piece stays with the two that cause almost all the disagreement: sharpness and eyes open. The third troublemaker, near identical frames, has its own piece.

Blur and blink cause arguments for the same reason. Both measure something real. Neither measures the question in your head.

You want to know whether focus landed on the bride's eye. The software knows how fast brightness changes across edges somewhere in the frame. You want to know whether she looks present. The software knows how far apart her eyelids are. Everything that goes wrong lives in that gap.

What follows is the mechanism, then the catalogue of frames that fool each check.

Blur is three problems sharing one number

Three different things get called blur, and only one is a mistake.

• Missed focus. The lens focused on the wrong plane. The nose is crisp, the eye is not.
• Camera shake. The whole frame smeared because the shutter stayed open longer than your hands stayed still.
• Subject motion. The frame is sharp and the moving thing in it is not. Sometimes an accident, often the point of the photo.

A sharpness score cannot tell these apart. It sweeps a small window across the frame, asks at every point how quickly brightness is changing across edges, and adds it up. All three problems produce the same symptom, slow change at the edges, so all three produce the same low number.

The score depends on where you point it

A whole frame score is dominated by whichever part of the picture has the most edges. Sequins, lace, foliage, brick. One calm face against a painted wall gives the measurement almost nothing to find, so it scores low while being perfectly sharp, and a frame that missed the face scores high because the beading on the dress is crisp.

Tools that find the face first and measure inside that box are far more useful. Narrative Select describes its focus assessment as telling you how well your subject is in focus, which is the right unit. Imagen describes its version as detecting out of focus shots. Check which one you are buying.

Grain reads as detail

High ISO noise is brightness changing fast from pixel to pixel, which is exactly what the measurement counts. A noisy frame from the dark end of the reception can outscore a clean one from the ceremony, and in camera noise reduction smooths grain away, so two of your bodies score the same scene differently.

It measures whatever it was handed

Many tools read the JPEG preview embedded in the RAW rather than the RAW itself, because it is far faster, and preview size varies by camera and by a menu setting most people never open. A small preview hides small blur. So a vendor's accuracy figure says little about your files. Run your own test instead.

A catalogue of frames that fool sharpness

The last row is the expensive one, because it is the failure you never catch. Everything above it gives you a score to argue with. Focus landing on the eyelashes instead of the iris gives you a score you agree with, until somebody orders a forty inch print.

One structural point. Inside a group of near identical frames, most tools rank rather than judge: sharpest of the six, not whether any of the six is sharp enough. If autofocus missed the whole run, you are still handed a best in group, and it is still soft.

How a blink becomes a number

Eye detection runs in two steps, and the order tells you where it breaks. Step one, find the face: a detector returns boxes plus a confidence value. Step two, look inside the box: landmark points go on the corners of the eyes and along the lid lines, and the software works out how far apart the lids are compared with the width of the eye. Wide gap, open. No gap, closed.

The quiet failure is step one

If no face is found there is no eye score at all. The frame is not marked closed, it is judged on sharpness alone, and in most interfaces that looks exactly like a frame that passed. Profiles, backs of heads, a guest thirty feet away in dim light. Not a wrong answer, no answer. Learn where your tool shows the no face pile.

A blink is a slope, not a switch

Blinking is continuous movement. Fire at eight to ten frames a second and you catch lids at every position on the way down and back up. Half lidded frames land in the middle of the range, precisely where the threshold sits, so tiny differences flip the verdict. Some half lidded frames look relaxed and lovely. Others look like the best man had a long night. The measurement cannot tell those apart, because the difference is in the mouth and the eyebrows.

The ratio is calibrated on other people's faces

Lid gap divided by eye width is a ratio, and eye shape varies enormously between people. Someone with naturally narrow eyes or a heavy lid can produce a fully open ratio sitting near the line the software drew from its training examples. Older subjects and sleepy toddlers push the same way. The number is real. What it means is not the same on every face in the room.

A catalogue of frames that fool eye detection

That last row deserves as much attention as the rest of the table together. Everyone worries about frames the software wrongly flags. The frames it wrongly passes are the ones that reach a client gallery, because nothing asked you to look again.

The vendors are open about the flagging side. Imagen ships a kiss recognition exception on the grounds that some closed eyes are still keepers, a carve out that exists because the eyelid measurement gets kisses wrong. Narrative reports eye state and lists blinking, kissing and looking down together, a measurement handed to you rather than a verdict handed down.

The group photo problem

Everything above assumes one face. A formal has twelve, and this is where the eye check stops being a measurement and becomes an editorial decision made on your behalf.

The software produces twelve eye scores, then has to turn them into one number for the frame. Three common ways to do that, and they disagree:

• Worst face wins. One cousin in the back row blinks, and the frame where the couple look wonderful sinks below one where everybody looks blank.
• Average the faces. One closed pair gets diluted by eleven open ones and disappears. You find it during editing.
• Count the closed ones. Now you have a threshold on top of a threshold, deciding how many blinks are acceptable before you have seen the photo.

Most tools do not say which they use. So stop trusting the frame level number on formals. Work at the group level: look at every frame of the setup together, check faces rather than scores, and expect the winner to come from the middle of the run. More in our guide to culling group photos for open eyes.

Six habits that keep both checks honest

• Sort by the score, never filter on it. Sorting reorders your work. Filtering hides part of it, and the moment frames are hidden, the threshold has quietly become the decision maker.
• Open the no face pile first. Those were judged on sharpness alone: profiles, backs, distant guests, anyone in deep shadow, and a lot of your reportage.
• Work the middle of the range, not the ends. Confident scores at either end are the easy cases. The middle is where half lidded frames, narrow eyes, glasses and crying pile up.
• Zoom to the eye on anything going out large. No sharpness score knows whether focus landed on the iris or the lashes. That check belongs on a big screen, which is why we treat phone and computer culling as two passes.
• Mark deliberate blur before you run anything. Flag the veil toss and the drag shutter set at ingest, or spend the evening rescuing them from the bottom of a sorted list.
• Test on your own work. Run a wedding you already picked by hand and count the disagreements both ways. That number beats any published accuracy claim, as we argue in can you trust AI culling.

Underneath all six is one question: how expensive is it to disagree? If seeing a frame it passed over takes four clicks and a settings change, you stop looking, and it starts making your editorial calls by attrition. One tap, and it stays an assistant. Which is also why we wrote about the photos you do not pick.

How we are approaching this at Kepla

We are building a picking app, so read this as interested rather than neutral.

Two decisions follow from everything above. First, every pick carries the reason it made the list: sharpest of its group, eyes open, the only clean frame of that moment. A reason is arguable in a way a bare score is not. See sharpest in group on a frame where you wanted the softer one, and you spot the disagreement instantly.

Second, nothing is ever deleted, moved or renamed. Kepla only marks which photos it picked. Every original stays where it was, and every frame it passed over is one tap from being added back. Given how easily both checks are fooled, a tool that files your alternates elsewhere is asking for trust it has not earned.

Status, plainly: the picking app for iPhone, iPad and Mac is still being built. What is live today is the Kepla booking page, free while we build.

FAQ