Kalal's Predator tracker learned from its own misses to follow any object

Computer-vision object tracking in 2011 usually meant following a fixed appearance for a short stretch: move the camera, turn the object, or let it leave the frame and the tracker lost it. What demos did not offer was a tracker that could follow almost anything a camera saw once it was selected, and that kept getting better as it ran.

Zdenek Kalal, a Czech PhD student at the University of Surrey in the United Kingdom, built Predator as part of his doctorate and demonstrated it on video: he selected his own finger once, and the system then tracked his hands almost perfectly, learning from the movement so that tracking improved over the session. A short training pass, the write-up noted, could teach the same system to follow a fingertip, a face, or a car racing down a highway.

The underlying choice was to let the tracker learn while it was being used rather than fixing it at build time. The demonstration was framed as an algorithm that learns from mistakes, and a series of twelve videos plus code posted at the university website let anyone watch the behavior and test the claim for themselves.

The article read the result as a path to hands-free interfaces: with the algorithm built into an operating system, a dot on the forehead, a fingertip or a moving head could steer a computer, and people without the use of their hands could gain an interface. When the demo reached Hacker News on April 3, 2011, it drew 460 points and 76 comments, and Kalal was recognized with a "Technology Everywhere" award.

One click seeds a tracker; every miss trains it live

Learning happens during use, so accuracy compounds while the video plays instead of degrading as the object changes.

One click defines the target, so the same algorithm can follow a finger, a face or a car without building a new model for each object.

Because it learns from what the camera shows, including its own misses, a short training pass can adapt it to changes in scale, orientation and visibility.

Publishing demo videos and code turned a PhD project into something thousands of engineers could watch and probe, which is what carried the idea.

When a system must handle objects it was never pre-trained on, let it learn from its own runtime misses instead of shipping a frozen model: live feedback turns each failure into the next improvement.

Predator reached a technical audience in April 2011, drawing 460 points and 76 comments on Hacker News. Readers found Kalal's page at the University of Surrey, noted the papers behind the method, and saw that the code, advertised as freely available, was then released only on request by email under a GPL notice, which several commenters challenged. Others probed where it would break, such as a fingertip lost behind a hand. The article records a "Technology Everywhere" award and foresees interfaces steered by a moving head; no commercial product or measured adoption appears in these sources.