search

UX-Decoder / DINOv

[CVPR 2024] Official implementation of the paper "Visual In-context Learning"
321 stars 11 forks source link

Problem with panoptic segmentation for ADE20k #15

Open agnieszkalysak opened 3 months ago

agnieszkalysak commented 3 months ago

Hi, I am trying to reproduce your evaluation on _"ade20k_panopticval" dataset, but there is issue with the visual prompts generation. The script is looking for anns[ann['id']] = ann and giving KeyError: 'id'.

As I see, the Detectron2 panoptic datasets doesn't have id field natively, the data was preprocessed accordingly to the MaskDINO instruction, which you provided, and everything worked fine there.

Please, could you explain more about how visual prompts could be generated from this data? Thank you in advance :)

FengLi-ust commented 3 months ago

Hi, I rerun and did not see a problem in ADE20K evaluation. Can you explain your questions with more information?

agnieszkalysak commented 3 months ago

Thank you very much for your reply. I solved the problem with id, the issue was related to the fact, that I was registering dataset to the Detectron2 catalogue, not using your scripts from datasets/registration, but my own method, from Detectron2 Docs. With your registration, everything worked perfectly.

But still, the metrics have rather low values - I have problem with achieving similar results to yours.