Inference on an image

[Vicktoria13]

Hello,
Thanks for your work ! I would like to perform an inference on my own images. I used detectron2 and the forward function of your model class EBSeg() , and looked the output : it gives me an array of shape (171,H,W), and when doing the max over the first dimensions to get an id segmentation map, I get only one class represented ... There is something I don't do correctly, hence my questions :

1. Do you provide any script that do a zero-shot image inference and visualization like in your paper ?
2. Also, do we have to provide out set of classes along with the image ?
3. in the forward function of the model, there are those lines :
   dataset_overlap_map = self.dataset_overlap_maps[dataset_names[0]].to(self.device)

why does the parameter dataset_names have an impact on the calcul ?

Thanks !