search

ksaito-ut / openworld_ldet

MIT License
54 stars 2 forks source link

agnostic setting #1

Closed wangchust closed 2 years ago

wangchust commented 2 years ago

Hi, I have two simple questions.

  1. How to set config file when train plain maskrcnn in an agnostic way.
  2. why set num_classes to 80 under VOC to NonVOC setting.
wangchust commented 2 years ago

By the way, the link of maskrcnn trained weights seems to be ldet model. Because it get the same scores as ldet gets in paper

ksaito-ut commented 2 years ago

Thanks for reporting the issue on the mask rcnn trained model. I will look into the issue.

  1. First, you should set the number of class as 1 in your config. Also, in the training script, i.e., trainer_plain.py, you need to set the classes of instances as 1, like for da in data: da['instances'].gt_classes = da['instances'].gt_classes * 0 in line 141.

  2. It is just because we kept the class index of COCO. We just avoided mapping the classes to that of 20. I guess it will not make a large difference in the performance.

wangchust commented 2 years ago

Thank you for your reply. I'll close this issue