DiffusionDet — Motion Detection With One Training

At last a system that improves motion detection with number of boxes and iterations rather than by retraining the model!

Motion detection is used by autonomous cars and home security systems.

Inevitably such systems became our digital pugs, protecting us from threads and supporting us in our daily lives.

Scientists still research ways to improve such systems. Making them detect an intruder every time, differentiating a human from a cat, and doing it as fast as possible.

There is a broad field for improvement. Fast and accurate motion detection can prevent accidents. Production line systems can detect bad products before they land in our homes.

How motion is recognized?

Motion detection systems are complicated, but usually contain a machine learned neural network.

Scientists grab a public set of training data, for example MS COCO (a set of “tests” with answers included), train the system on these data, and than, run the system on a validation set, to check if the system “guesses” the answers right.

That way, scientists creating various motion detection systems can compare results with…