search

cleinc / bts

From Big to Small: Multi-Scale Local Planar Guidance for Monocular Depth Estimation
GNU General Public License v3.0
635 stars 179 forks source link

How to deal the invalid area of the depth? #114

Open mathmax12 opened 3 years ago

mathmax12 commented 3 years ago

Thanks for the amazing work and code.

When checking the depth map there are some invalid areas. I wonder how you deal with this part during training, validation, and test.

Thanks.

cogaplex-bts commented 3 years ago

@mathmax12 We simply ignore invalid pixels that do not have depth values. Please check our implementation. Thanks.

mathmax12 commented 3 years ago

@cogaplex-bts Thanks. I got it. Another question. in the paper "Because weights at early convolutions are known to be well trained for primitive visual features, in the base networks, we fix the first two convolutional layers as well as batch normalization parameters in our training." How to choose how much the first layers should we keep fixed? In the paper, I saw the first two, and the batch norm is fixed.

Thanks

mathmax12 commented 3 years ago

Another question is when using the nyu to training the model, for the loss calculation, the mask is generated with if args.dataset == 'nyu': mask = depth_gt > 0.1 I wonder how chose the 0.1 here. Thanks.