joyhsu0504 / darcnn

MIT License
5 stars 1 forks source link

Reproducing doubts #2

Open Xujingkk opened 3 months ago

Xujingkk commented 3 months ago

Hi, Joy Hsu, thanks for the interesting paper and open sourcing the code here. I'm trying to reproduce the code to understand more details. But I have some questions I would like your help with:

  1. I saw how to use DARCNN at "darcnn_code/tutorial.ipynb" and the task demonstrated is the adaptation from coco to BBBC, right?
  2. If I want to try adaptation from BBBC to Kumar, then in the first stage I need to load BBBC dataset and Kumar, and in the second stage only Kumar, right?
  3. When adaptating from BBBC to Kumar, can I still use the pre-trained weights 'coco_class_agnostic_maskrcnn.pth' in the first stage? If not, what should be done here?
  4. I downloaded the bbbc_10k_256.zip that you shared, is this data already done with all preprocessing including inversion? If I have misunderstood, I would appreciate your clearer guidance!
Xujingkk commented 5 hours ago

Hi, Joy Hsu, Regarding kumar's data preprocessing, are there 10,000 randomly cropped patches from the official 16 training images? Is there any other data augmentation done?