Pixel mean used for Resnet152

MILVLG / bottom-up-attention.pytorch

A PyTorch reimplementation of bottom-up-attention models

Apache License 2.0

294 stars 76 forks source link

Pixel mean used for Resnet152 #79

Closed SkrighYZ closed 3 years ago

SkrighYZ commented 3 years ago

Hi,

I want to extract features on custom datasets using the ResNet152 model. The pixel mean stated in the extract-bua-caffe-r152.yaml file is [0, 0, 0]. Should I change it to [102.9801, 115.9465, 122.7717] as in the ResNet101 case, or some other value?

Thanks in advance.

MIL-VLG commented 3 years ago

The mean [0,0,0] is only for R-152 backbone. If you use R-101, then you should use [102.9801, 115.9465, 122.7717]