Why not continue with vector attention?

Pointcept / PointTransformerV3

[CVPR'24 Oral] Official repository of Point Transformer V3 (PTv3)

MIT License

728 stars 42 forks source link

Why not continue with vector attention? #57

Closed lmomoy closed 3 months ago

lmomoy commented 3 months ago

I notice that the PT v2 uses vector attention like v1, while PT v3 uses scalar attention. Why not continue with vector attention?

Gofinge commented 3 months ago

Hi, it is a good question; for details, please read our paper Sec 4.2 Paragraph 1: Re-weigh options of attention mechanism. In one word, vector attention is for unstructured data, and scalar attention is for structured data.