search

shekkizh / FCN.tensorflow

Tensorflow implementation of Fully Convolutional Networks for Semantic Segmentation (http://fcn.berkeleyvision.org)
MIT License
1.25k stars 527 forks source link

Why the scale factor in caffemodel is ignored in tf equivalent? Should we scale the input to 1/256 or 1/128 as caffe used to do? #48

Closed Edwardmark closed 7 years ago

Edwardmark commented 7 years ago

Why the scale factor in caffemodel is ignored in tf equivalent? Should we scale the input to 1/256 or 1/128 as caffe used to do?

Edwardmark commented 7 years ago

Why not scale the image to have maximum of 1?

shekkizh commented 7 years ago

The pre-trained vgg model afaik used only mean subtracted images without scaling. Since we initialize with trained values I think it's fair to not scale the images.

Can you point to where you found scaling using in models trained in caffe?