Hi,
I am try to reproduce the deeplab results using VOC2012 data set as described in instructions at
https://bitbucket.org/aquariusjay/deeplab-public-ver2
I tried two approaches 1) using already trained model vgg16_20M.caffemodel and created a symoblic link of this model with name init.caffemodel in deeplab_largeFOV directory and fine tuned it secondly I tried to train without using init.caffemode. In first case training is going fine and prediction accuracies are good. But when i use second approach, i observe that only background class is predicted and test, evaluation phases show that only background class has some accuracy whereas all other classes are predicted to be zero.
Can you please guide in this context.
Best
Hi, I am try to reproduce the deeplab results using VOC2012 data set as described in instructions at https://bitbucket.org/aquariusjay/deeplab-public-ver2 I tried two approaches 1) using already trained model vgg16_20M.caffemodel and created a symoblic link of this model with name init.caffemodel in deeplab_largeFOV directory and fine tuned it secondly I tried to train without using init.caffemode. In first case training is going fine and prediction accuracies are good. But when i use second approach, i observe that only background class is predicted and test, evaluation phases show that only background class has some accuracy whereas all other classes are predicted to be zero. Can you please guide in this context. Best