search

HHHedo / IBMIL

CVPR 2023 Highlight
72 stars 9 forks source link

About pre-computed features #2

Closed jjjharden closed 1 year ago

jjjharden commented 1 year ago

Hi, I am trying to implement your code.The input image size of CTransPath is 224×224, but you mentioned in your paper that the image size you used is 256×256, how did you handle it when calculating the features?

HHHedo commented 1 year ago

Hi, we just use transforms.Resize(224) in computing features.

jjjharden commented 1 year ago

1 Hi, here is the pre-computed features I obtained using Ctranspath, but the result of training the MIL aggregator with this pre-computed features is very poor.Is the data for this pre-computed features correct?

HHHedo commented 1 year ago

1 Hi, here is the pre-computed features I obtained using Ctranspath, but the result of training the MIL aggregator with this pre-computed features is very poor.Is the data for this pre-computed features correct?

Hi, the effectiveness of CTransPath has been validated in many papers. You may directly leave an issue in CTransPath's official github. For your features, I empirically sample some slides and find the features are around [-1,1], but your features are near 0 (which may be caused by random initialization). Please make sure you correctly load the pre-trained weights and extract the features following dsmil.

jjjharden commented 1 year ago

Thanks a lot!