search

scaelles / DEXTR-PyTorch

Deep Extreme Cut http://www.vision.ee.ethz.ch/~cvlsegmentation/dextr
GNU General Public License v3.0
843 stars 153 forks source link

pretrain the basenet #12

Closed JongMokKim closed 6 years ago

JongMokKim commented 6 years ago

Hi, thank you for nice work!

I have a question about pretraining the basenet on imagenet

your base network has 4 channel input (h x w x 4(rgb-guassian)), but it should be 3 channels when pretraining the imagenet (imagenet data only has rgb and extreme points is not necessary to classify task)

how did you do it? thank you for your help in advance!

kmaninis commented 6 years ago

Hi, thanks for your kind words.

The network was indeed pre-trained on Imagenet using 3 channels. We artificially added a fourth channel to train DEXTR, and initialized its weights from Imagenet. Please take a look at this part of the code.

JongMokKim commented 6 years ago

thank you very much!