search

microsoft / USBuildingFootprints

Computer generated building footprints for the United States
Other
2.12k stars 264 forks source link

training dataset information #50

Open amardeepjaiman opened 5 years ago

amardeepjaiman commented 5 years ago

Hello,

Could you please confirm if you use the labelled training data of same geographic location (US) to train your FCN network ? Second, as you mentioned that the training data was having resolution of 1ft/pixel, do you have to use same resolution images for predictions also or this trained work is able to generate building footprints on any resolution image ?

Thanks, Amardeep

nitrif commented 5 years ago

Hi Amardeep, Yes we used only US labels in this project and the model was trained only on 1ft/pixel images. We have never tried inference on different resolutions, but the expectation is that it won't work good. Regards, Nikola

amardeepjaiman commented 5 years ago

Thanks Nikola for taking out time to reply. Same expectation i have that it shouldnt work on other resolutions. could you share the network architecture information you are using for training, ?

nitrif commented 5 years ago

It is a standard ResNet with RefineNet upsampling.

amardeepjaiman commented 5 years ago

ok. Can you guide how we can generate separate masks for same class like instance segmentation using RefineNet as it is for Semantic segmentation like task. isnt it ? Is it fine If we use Mask RCNN for this task which is specifically for Instance Segmentation. do you think any benefit of using RefineNet over Mask RCNN ?