Inside the Hyperspectral Object Tracking Challenge 2026 | by Alan Scott Encinas | Jun, 2026 | MediumSitemapOpen in appSign up<br>Sign in
Medium Logo
Get app<br>Write
Search
Sign up<br>Sign in
Inside the Hyperspectral Object Tracking Challenge 2026
Alan Scott Encinas
4 min read·<br>3 hours ago
Listen
Share
In the movie, the Predator does not see the jungle the way you do. It reads heat, then flips to another band, hunting a signature the human eye throws away. I have spent the last few weeks building something with the same job. The difference is that my target is real, and the footage comes from a sensor that records colors your eyes were never built to receive.<br>This is the first entry in a builder’s log I’m keeping while I actually compete. No tidy after-the-fact write-up where I already know how it ends. You get it as it happens.<br>What this actually is<br>The competition is the Hyperspectral Object Tracking Challenge 2026, run on Kaggle alongside WHISPERS, an academic conference on hyperspectral imaging. The task sounds simple and refuses to be. You are handed a single object, boxed in the first frame of a video, and you have to find that same object in every frame after it. Frame one, here is the target. Frame two through the end, where did it go.<br>That problem, single-object tracking, is old and well-studied for normal video. People have been chasing it since before deep learning was cool. What makes this one different is the camera.<br>The camera sees colors you don’t<br>A normal video gives you three numbers per pixel: red, green, blue. Your phone does it, your eyes do it, every movie you have ever watched does it. A hyperspectral video gives you sixteen numbers per pixel. Or twenty-five. Each one is a thin slice of the spectrum, and some of those slices sit past the edge of what a human retina can register.<br>This is the whole game, and it cuts both ways. A target and its background can look identical to you and be obvious in a band you have never seen. A tank under a camouflage net, a specific mineral on a hillside, one particular boat in a marina. The information is there. It is just hiding in a channel your eyes skip.<br>The footage comes from three different sensor families, and they do not behave the same way. Some carry sixteen bands, some twenty-five, some fifteen, and what is easy in one is brutal in another. You are graded across all of them at once, which means there is nowhere to hide a weakness.<br>The stakes and the clock<br>This competition has a two-sided prize, and that is what pulled me in. The top of the Kaggle leaderboard is a podium. The top four teams also have to write and submit a paper to WHISPERS. So this is one of those rare contests where winning and publishing are the same act. You do not get the medal without the research.<br>The clock is real. The final test data drops on September 7, submissions close on September 10, winners are announced September 15. That gives a runway measured in months, not weekends.<br>You are scored on two things at the same time: how tightly your predicted box overlaps the real one, and how close the center of your box stays to the target as it moves. One number rewards precision of shape. The other rewards never losing the thing. A tracker can be good at one and quietly terrible at the other, and the score will find you out. The podium line sits around 0.56. The leaders are near 0.66. The floor, if you basically do nothing, is a rounding error above zero.<br>How I’m going to write this<br>Here is the part that makes this a log and not a press release. I am writing these while the competition is live, so I have to be careful. I publish from behind a delay. You will see where I was, not where I am. The moves that are still my edge stay dark until they stop being an edge. Everything else, the dead ends, the gotchas, the bug that cost me a full day, you get all of it.<br>And I am not running only this one. There are seven of these going at once. That is the only reason I can write honestly without handing a competitor my playbook. There is always something far enough behind the front line to be safe to tell.<br>Where I’m standing right now<br>I have a baseline on the board at 0.524.<br>That is real, and it is about three and a half hundredths under the podium line. Close enough to be maddening. Far enough to be the entire project. Everything from here is the work of closing that gap, and a competition like this is usually won or lost in the last tenth of it.<br>Tracking a target through light we cannot see sounds like science fiction until you notice that a surveillance satellite, a search-and-rescue drone, and a medical scanner all want the exact same thing. The Predator’s party trick turns out to be a serious engineering problem with serious people working on it. I would like to beat some of them.<br>Next entry: the first wall I hit, and it was not the modeling. It was the discovery that the leaderboard I am chasing cannot actually tell me my score.<br>This is part of an ongoing builder’s log written...