Decoupled Pipeline with Proposal Reranking and Score Fusion for Positive-Unlabeled Marine Species Detection

Imagine you are a detective trying to solve a mystery in a giant, murky underwater city. Your job is to find specific sea creatures and name them. But there's a catch: the map you were given is full of holes. The people who drew the map only wrote down the animals they were sure about, leaving thousands of other fish, urchins, and jellyfish unmarked. In the world of computer science, this is called "positive-unlabeled" learning. It's like trying to learn what a "dog" looks like by only studying pictures where someone explicitly pointed at a dog, while ignoring all the other animals in the photo that might just be cats or squirrels. If your computer sees a cat and thinks, "That's a dog!" because it wasn't told otherwise, it gets confused.

To make things even trickier, the underwater city is a bit like a shifting dream. The photos you are studying were taken by one group of divers in one part of the ocean, but the "test" photos you need to solve are from a completely different ocean, taken with different cameras and lighting. This is called "domain shift." It's like trying to recognize a friend's face after only seeing them in black-and-white photos, but then being asked to identify them in a high-definition, color video. The goal of this research is to build a computer system that can spot these sea creatures and name them correctly, even when the map is incomplete and the lighting is weird.

---

The Detective's New Strategy: Splitting the Job

The team from Georgia Tech, known as DS@GT ARC, realized that trying to do everything at once was a recipe for disaster. Imagine asking a single person to be the eyes that spot the animal, the brain that names it, and the judge that decides if the guess is good enough. In their experiments, this "all-in-one" approach hit a wall. The computer kept getting confused by the unmarked animals, thinking they were mistakes.

So, they decided to split the job into a relay race with three distinct runners, each doing one thing very well.

Runner 1: The Wide-Eyed Scout (The Detector)
First, they needed someone to scan the whole ocean floor and shout, "Hey, there's something there!" They used a pre-trained robot called "Megalodon" (a type of YOLO detector). Instead of teaching Megalodon to learn new things from the messy competition data, they kept it frozen, like a veteran scout who already knows how to spot shapes.

• The Tiling Trick: The ocean floor is huge, and some sea creatures are tiny. If you squint at a whole photo, a tiny urchin looks like a speck of dust. To fix this, the team didn't just look at the whole picture; they sliced the image into smaller, overlapping tiles, like a mosaic. This let the scout zoom in on the tiny details.
• The Edge Filter: Sometimes, a creature gets cut in half by the edge of a tile. The team added a rule: if a detected shape is chopped off by the tile edge, throw it away. This kept the team from counting half-urchins as whole ones.

Runner 2: The Expert Biologist (The Classifier)
Once the scout found a "something," they passed a cropped photo of that spot to the second runner: a super-smart AI called DINOv3. This runner's only job is to look at the creature and say, "That's a sea urchin," or "That's a bony fish."

• The Secret Sauce: They didn't teach this robot from scratch. Instead, they used a technique called "LoRA" (Low-Rank Adaptation). Think of this as giving the robot a small, specialized notebook to write down the specific rules for the 32 types of sea creatures in the competition, without rewriting its entire brain. This allowed the robot to become a master of fine-grained details, like telling the difference between two very similar types of fish.

Runner 3: The Doubting Judge (The Reranker)
Even the best biologist can be overconfident. Sometimes, a weird rock might look enough like a fish that the biologist says, "99% sure that's a fish!" The team added a third step to check the confidence. They built a "validity head" inspired by a "Turing Test" (a test to see if a machine can fool a human). This judge asks: "Does this candidate actually look like the other members of its class?"

• The Result: This judge acted like a gentle nudge. It didn't throw out guesses; it just slightly adjusted the ranking. If a guess was weird, it got pushed down the list. If it looked solid, it stayed near the top.

The Final Score: Putting It All Together

The team didn't just let the three runners shout their answers. They used a special math trick called "weighted geometric fusion" to combine the scores. Imagine the Scout gives a score of how sure they are that something is there, the Biologist gives a score of how sure they are of the name, and the Judge gives a score of how consistent the guess is. The final score is a blend of all three. If any one of them is weak, the final score drops, ensuring only the strongest, most consistent guesses make it to the top.

What They Found (and What They Tossed)

The team ran hundreds of experiments, and the results were a mix of "aha!" moments and "oops" lessons.

• The Frozen Scout Won: They tried teaching the scout (Megalodon) to learn from the messy data, hoping it would get better. But it didn't work. The "frozen" scout, which just used its existing knowledge, actually performed better in the final race. The team learned that trying to fine-tune the detector on noisy, incomplete data made the whole system worse.
• The "Strict" vs. "Real" Trap: They tried to be super strict, only accepting guesses that perfectly matched the known labels. But in the real world (the hidden test set), many real creatures weren't labeled in the training data. If they filtered too aggressively, they accidentally threw away real animals that just happened to be unmarked. The lesson? It's better to keep a wide net and let the ranking system sort the good from the bad, rather than deleting anything that looks suspicious.
• The Public vs. Private Split: The team had two versions of their system. One version, which included the "Doubting Judge" (the TTN-inspired reranker), did great on the public leaderboard and their own external tests. However, on the final, secret private leaderboard, the simpler version (without the judge) won by a tiny margin. This taught them that a signal that looks helpful in one test might not be robust enough for the final, unknown challenge.

The Big Picture

In the end, the team's system placed 12th out of 102 teams with a score of 0.1757. Their biggest discovery wasn't a new super-algorithm, but a strategy: don't try to fix the messy data; work around it.

They found that separating the job of "finding" from the job of "naming," using a frozen detector, and relying on a smart ranking system was more powerful than trying to force a single model to learn everything. They also learned that in a world of incomplete maps, being too aggressive about filtering out "doubtful" guesses is dangerous. Sometimes, the thing you think is a mistake is actually a real, unmarked treasure.

The paper suggests that while using "student-teacher" methods (where a smart model teaches a weaker one) and creating fake data are promising ideas, they need to be handled with extreme care. If the teacher isn't perfect, the student just learns the teacher's mistakes. For now, the safest bet is to keep the detector simple, let the classifier do the heavy lifting on names, and use a gentle hand to rank the final guesses.