search

juhongm999 / hsnet

Official PyTorch Implementation of Hypercorrelation Squeeze for Few-Shot Segmentation, ICCV 2021
231 stars 43 forks source link

backbone feature #27

Closed Catofblue closed 2 years ago

Catofblue commented 2 years ago

thanks you nice work ! I have a question, you use the backbone of resnet 50 and resnet101 to extract feature, why use 4 to 16 layers features

res50 feat_ids = list(range(4, 17))

    if hid + 1 in feat_ids: 
        feats.append(feat.clone())

    feat = backbone.__getattr__('layer%d' % lid)[bid].relu.forward(feat)

return feats
juhongm999 commented 2 years ago

We use feature layers of stages 2, 3, and 4 of ResNet which correspond to those layer indices.