AI’s imagination is terrifying: it can match 5 different full-body poses just by getting a part of the human body

This is a three-dimensional representation that encodes surface shapes as discrete sequences of two-tuples, each representing the location and content of a local feature, which is a process like this:
Second, the authors propose a Transformer-based autoregressive model, ShapeFormer, which sequentially predicts the distribution of the complete sequence based on the sequence of bigrams generated in the previous step.

The Transformer here is a model in 2014, which can use the attention mechanism (Attention) to improve the training speed of the model. It has made breakthroughs in the field of natural language understanding (NLP) since it was first launched, and many have applied it in recent years. Cross-border research in the field of computer vision (CV).

In addition to flat shapes like tables and chairs, AI has also learned many skills such as symmetry, hollowing, and filling during training, so it can also generate 3D models such as teapots or kettles.

At the same time, the “imagination” of this kind of AI is also extremely rich. For example, in the face of a human pose with a generation target with great uncertainty, the author randomly selects a small part from the complete model, while the AI keeps observing the body Under the premise of the pose of the part, a variety of possible poses are also generated.