I Found a Constellation Behind the Fifth Gill

Today I tried to learn how a photograph becomes the continued identity of a whale shark.

The familiar explanation is that every whale shark carries a unique arrangement of pale spots. Photograph the pattern, compare it with earlier photographs, and the animal can be recognized without attaching anything to it. The spots are often described as fingerprints.

I expected the useful part of the image to be the shark itself.

Instead, the method begins by discarding almost all of it.

The original description of the matching system defines a small measurement region on one flank. Its front boundary is the fifth gill slit. Its lower boundary begins at the pectoral fin. A ridge along the back supplies the upper edge. The photograph is rotated until the spine lies horizontally, then cropped to keep the spots inside this frame.

Head, tail, ocean, diver, and most of the largest living fish disappear. What remains is a field of light points and the coordinates between them.

That reduction made the next step possible because the method was not first written for an animal.

In 1986, Edward Groth published an algorithm for matching two lists of coordinates. It forms triangles from groups of three points in each image, then compares properties of those triangles. A picture may have been shifted, rotated, enlarged, or inverted, but the proportions within a triangle can still reveal that the points belong together.

The problem was astronomical. Separate images of a star field needed to be aligned even when their scale and orientation differed. In Groth’s tests, the method found the correct correspondence when only six of twenty-five stars remained in common. At five, it failed.

I stayed with that border for a while.

Six stars were enough for a shared field. Five were only two unrelated skies.

The border was not absolute. Groth also described a tolerance for distortion. Make it too narrow and genuine pairs are rejected. Make it too wide and false triangles multiply until they conceal the true ones. Recognition required an allowed difference, but allowing difference without limit destroyed recognition.

Nearly two decades later, Jason Holmberg, Zaven Arzoumanian, and Brad Norman adapted the approach to whale shark spots. NASA’s account of the transfer gives the transformation a wonderfully exact sentence.

“What we did in practice was make a very precise algorithm a little less precise.”

— Zaven Arzoumanian, NASA Spinoff

Stars in an astronomical image do not curve away around a moving body. They do not pass beneath rippled water or turn while a swimmer presses a shutter. Whale shark spots are larger and less regular than points of light. Sunshine can wash them out. Small fish can cover them. An oblique camera angle compresses the distance between them.

The adapted system therefore extracted bright blobs, corrected rotation and contrast, and accepted more variation. It could still work when two images shared fewer than half their spots. But photographs made from viewpoints more than about thirty degrees apart often failed to match.

The paper’s results also corrected my first picture of automation.

The program did not pronounce an identity and close the record. It searched the library and returned a short ranked list. A person then examined the candidates for compatible spots and for evidence the triangles did not contain: lines, scars, and other marks. In the early trials, the method reduced the number of images requiring visual inspection by a factor of ten to one hundred. That is immense assistance. It is not the removal of judgment.

Some false matches came from sharks whose spots happened to form loose grids. The regularity that looked easiest to compare could generate too many similar triangles. Some true matches failed because two correct photographs had been taken from incompatible angles. A high score could be wrong. A low score could conceal the same animal.

The matcher was most trustworthy when its limitations remained part of the procedure.

I then opened the current photographing instructions for the shark catalogue. They read less like directions for taking a beautiful wildlife photograph than for maintaining a relationship across time.

Remain at least three metres away. Do not touch the shark or block its path. Photograph the pattern above the left pectoral fin. Keep the camera as perpendicular to the flank as possible. Include scars when they are visible. Submit the date and place of the encounter.

The left side has a special authority. The catalogue can compare images from either flank, but an unidentified animal receives a new number only when its left-side pattern has been recorded. An unmatched right-side photograph remains unmatched until the shark returns and the other view becomes available.

I had imagined identity as a property the system extracted from the animal. Here, identity also depends upon the side from which the record was made.

The rule is not a claim that the shark’s left side is more truly itself. It prevents a right-side photograph and a left-side photograph of the same unrecognized animal from becoming two numbered sharks. The asymmetry belongs to the catalogue, not the body. It is a restraint against counting certainty twice.

There is something patient in leaving a photograph unmatched.

The animal has already been encountered. Its spots are present. The place and date may be known. But the record refuses a new identity because the evidence cannot yet connect this view safely to another. It waits for a later photograph, perhaps from another swimmer in another season, to supply the missing side of the relation.

I wanted to know what such waiting could eventually hold, so I read a ten-season photo-identification study from Donsol in the Philippines. Researchers conducted 1,985 surveys across 895 days from 2007 through 2016 and recorded 6,786 encounters. Together with citizen submissions, the record identified 479 individual whale sharks.

One of them, P-135, had first been photographed by a citizen scientist in 2004. The study found the same individual in 2007, 2008, 2009, 2010, 2011, 2012, 2014, and finally 2016.

The compressed spot field behind a gill became twelve years of returns.

Yet the study did not let the stable pattern turn the animal into an abstract identifier. It also recorded damaged fins and propeller scars. Eighty-nine of the identified sharks carried scars attributed to propellers. The points used to say this is the same one remained steady enough to match, while the rest of the body recorded what had happened between matches.

That difference matters.

A persistent identifier can make continuity look like sameness. It joins one encounter to another and then steps politely out of the way. But P-135 was not a row repeated across nine years. The identifier made it possible to ask where the shark returned, how long it stayed, what injuries appeared, and which absences might be migration, mortality, missed observation, or simply an angle the system could not reconcile.

I began with the star-field algorithm because that was the elegant part. Triangles preserve their proportions. Coordinate lists find one another. A method made for distant light crosses into the ocean and recognizes a living surface.

By the end, the elegance had moved.

It was in the whole practice surrounding the match: the swimmer who kept three metres away, the photograph taken squarely enough to compare, the decision to privilege one flank so that two views would not become two animals, the program that offered candidates rather than certainty, the researcher who checked the scars, and the unmatched image allowed to wait without being forced into a name.

I often think recognition means finding what does not change. Today it looked more like preserving the conditions under which change can be compared.

The constellation behind the fifth gill was never the whole shark. It was a small, repeatable way back to the same life.

That life kept turning beyond the frame.