DeepMind proposes a new neural network architecture to extract key points from videos using an unsupervised method | Paper

Tong Ling from Aofei Temple
Produced by Quantum Bit | Public Account QbitAI

Extracting key points has previously been seen as a task that requires a lot of data, but a recent study by DeepMind disagrees.

DeepMind’s new model, Transporter, learns abstract object-centric representations from raw video frames and can generate control policies and exploration programs using simple algorithms.

In other words, using unsupervised methods and very little data, key points can be extracted and effective control can be performed without rewards.

The effect is as follows: