jiajunwu / marrnet Goto Github PK

Hi,

Thanks for providing the codes. I am trying to re-train the second stage on shapenet data, but I am wondering what size you use when render the normal/depth, and create voxels. It seems they are all in the middle.

Thanks!

Hi @jiajunwu ,

Thanks for sharing the code.

I wonder if you can share the synthetic dataset that you mentioned in section4.1 and the pre-train code for both the 2.5 sketch estimator and 3D reconstruction model.

Hello, I am attempting to train marrnet on another dataset, but the depth I obtain from my dataset is very large and varies a lot in comparison to the shapenet dataset. Are the sizes of the meshes in the shapenet dataset constrained to a certain range?

Hi, @jiajunwu ,

Thanks for sharing your code. Would it be possible to have pytorch version of marrnet?

Please share the code for re-projection consistency based loss and optimization.

I want to ask you whether the demo_train_step2.lua includes fine-tuning process or not? Thank you very much .

Is MarrNet model trained on chair category only? I think the previous work 3D-R2N2 train the model on 13 ShapeNet categories? And DRC method is trained on three categories? Are the comparison fair in this case? Thank you!

Hi @jiajunwu,

Thanks for your work.

Could you share also the training dataset? input images, 2.5D sketches and voxels?
If they are already uploaded somewhere, could you indicate where? Thanks

To calculate the projected depth loss and projected surface normal loss, orthogonal projection is done (mentioned in 3.3). If so, then what is the shape of estimated depth and normal (2.5 sketch)? if the shape is not 128x128, then how orthogonal projection done from 128x128x128 voxel ?

Hi,

how long have you trained the models or how many epochs ?
Are the 2 steps trained the same amount of epochs?
How about the fine tuning process?

Thanks