search

scaelles / OSVOS-TensorFlow

One-Shot Video Object Segmentation
http://vision.ee.ethz.ch/~cvlsegmentation/osvos/
GNU General Public License v3.0
438 stars 132 forks source link

Why should make the deconvolutional part untrainable #9

Closed Beanocean closed 6 years ago

Beanocean commented 6 years ago

https://github.com/scaelles/OSVOS-TensorFlow/blob/f1eef4991547547c2a1a9fa720b3c832252b2617/osvos.py#L97

The code above shows the deconvolution layers are set to be untrainable, would you please help to explain this?

scaelles commented 6 years ago

We tried training the deconvolutional layers using the bilinear interpolation weights as initialization or directly training from scratch. The best result was obtained using the biliniear interpolation weights, but the performance was almost the same if we trained those layers or not, so we decided not train them.