crockwell / Cap3D

[NeurIPS 2023] Scalable 3D Captioning with Pretrained Models
https://huggingface.co/datasets/tiange/Cap3D
218 stars 13 forks source link

Scaling the CLIP Score #16

Closed AndreAmaduzzi closed 11 months ago

AndreAmaduzzi commented 11 months ago

Hello, I am wondering which is the meaning of the scaling by 2.5 which is applied to the CLIP Score in this file.

Thanks in advance, Andrea

crockwell commented 11 months ago

Hi Andrea,

Agreed this is a bit surprising practice, but it is standard practice. I think the reasoning is it causes values to approximately have 100 as max score.

Best, Chris

AndreAmaduzzi commented 11 months ago

I see. Thanks! I will take a look at that paper.

Best, Andrea