Detail about 'background' class

NVlabs / ODISE

Official PyTorch implementation of ODISE: Open-Vocabulary Panoptic Segmentation with Text-to-Image Diffusion Models [CVPR 2023 Highlight]

https://arxiv.org/abs/2303.04803

Other

845 stars 45 forks source link

Detail about 'background' class #37

Open develop-productivity opened 10 months ago

develop-productivity commented 10 months ago

Hi, thanks for your great work.

I have a question about the classifier in this paper, I want to know whether you used the 'background' class in the $C_{train}$.

We encode the names of all the categories in Ctrain with the frozen text encoder and define the set of embeddings of all the training categories' names as: Equation4

If you used a background class, is it learnable or fixed?