Affine
Innovation Lab · Competition Result

We Placed Top 10 of 256.
Then the Organisers Re-Ran Our Code.

9th of 256 teams in the Sixth Hyperspectral Object Tracking Challenge 2026, run in conjunction with IEEE WHISPERS.

Confirmed by an independent academic committee that took our code and tested it themselves.

Stacked spectral planes, each a different wavelength band, resolving one object
9th
of 256 teams
Top 4%
of the field
96%
of the winning score
2
rounds, including independent verification
The Challenge

A Research Problem, Not a Marketing Exercise

The Hyperspectral Object Tracking Challenge is in its sixth year. It runs alongside IEEE WHISPERS, the 16th Workshop on Hyperspectral Image and Signal Processing, held this November in Glasgow.

256 teams entered between May and September 2026. We came in as a team that ships AI into production, and finished in the top 4%.

Organised and sponsored by
  • Griffith University
  • Nanjing University of Science and Technology
  • Ghent University
  • Wuhan University
  • Helmholtz-Zentrum Dresden-Rossendorf
  • Grenoble Institute of Technology
IEEE WHISPERS 2026 · Glasgow · 17 to 19 November
In Plain Language

Seeing What the Camera Cannot

Standard video gives three values per pixel. Red, green, blue.

The cameras in this challenge captured 16, 25 and 15 bands, including wavelengths the human eye does not register. Every pixel carries a spectral signature rather than a colour.

The task was to lock onto a target and hold it across footage running at 25 frames per second. Teams trained against 406 hyperspectral videos, tuned on 75 more, and were scored on 75 they had never seen.

The targets were not convenient. They ranged from herbs and insects that disappear into the background to a motorcyclist at speed. The cameras were moving as well, mounted on drones and vehicles, so parked cars slide across the frame and a tracker will happily follow one instead of the target. Lose the thread for a few frames and it latches onto a wheel instead of the bike.

Two objects that look identical to an ordinary camera do not look identical in spectral data. That is the whole point, and it is also why the problem is hard.

A street scene with a motorcyclist, a drone and roadside plants each boxed as tracked targets, beside a strip showing the same scene captured at different wavelength bands
One scene, several bands. The targets a tracker has to hold, and the wavelengths an ordinary camera never records.
Dr. Param Jeet

Hyperspectral data does not arrive as a picture. It arrives as raw sensor grids that have to be rebuilt before anything can even look at them. That is the part nobody demos, and it is where most of the work is.

Dr. Param Jeet, VP - AI Practices, Affine
How It Was Judged

Anyone Can Top a Leaderboard. Fewer Survive the Review.

The challenge ran in two rounds.

ROUND 1

Open submissions scored against a held-out test set. 256 teams. The top ten advanced.

ROUND 2

The organisers took the advancing teams’ code, reviewed it, and re-ran it themselves.

That second round is the part worth paying attention to. Most competition results are self-reported scores on a public board. This one was reproduced by an independent committee before it was confirmed.

It is a different class of evidence from a case study written by the company that delivered the work.

Vineet Kumar

This is what 15 years of an uncompromising culture of innovation and exceptional talent can produce — deep technical expertise, original thinking, and the ability to solve problems that others struggle to crack.

Vineet Kumar, Chief Executive Officer, Affine
Why It Matters Commercially

The Same Problem, Wearing Different Clothes

This started as something our AI Centre of Excellence wanted to take on. It turns out to be the same problem as several things our clients pay us to solve.

The competition was the proving ground. The capability is what we ship.

The Team

The Same People Who Ship for Clients.

A team from our AI Centre of Excellence took this on alongside their delivery work, because the problem was worth solving.

They finished in the top 4% of 256 teams, and then had their work re-run by the organisers to prove it. We do not publish individual names. Internally, they know exactly what we think of this.

We are proud of this one, and of the team behind it.

That is the culture we hire for, and it is the reason results like this are not unusual here.

The Longer Record

Fifteen Years of Building It, Not Talking About It

Affine was founded in 2011, before the current category existed. Computer vision went into client production in 2016, and has been shipping ever since.

This result is the most recent entry in that record. It is not the beginning of one.

Talk to an Expert

Working on a vision problem that standard cameras cannot solve? Let’s talk about what production looks like.

Talk to an Expert