search

FilippoAleotti / mobilePydnet

Pydnet on mobile devices
Apache License 2.0
254 stars 40 forks source link

How does mobilePydnet architecture differ from Pydnet architecture #25

Closed zshn25 closed 4 years ago

zshn25 commented 4 years ago

You have mentioned in the paper that transposed convolutions have been replaced by upsampling and convolution blocks. What else has been changed?

FilippoAleotti commented 4 years ago

Hi, In the original pydnet, in each decoder we get the disparity by slicing the last feature map. Here we obtain disparities with a convolutional layer.

zshn25 commented 4 years ago

Thank you. Could you tell me why these changes? Why replace transpose convolutions and why convolution layer for disparity at the end of decoder?

FilippoAleotti commented 4 years ago

About transposed, also for this, while about disparity estimation, the number of parameters and time that slice allows to gain it’s quite limited, so I opted for these (minor) changes.

zshn25 commented 4 years ago

Off topic, but I've implemented both Pydnet and the mobile version in Pytorch here You can review if you have time and link it in the READMEs. Thanks!