yashkant/spad

custom dataset

Closed this issue · 2 comments

Excellent work!
I have a question: if I need to use my own dataset, how should I process it? Also, does depth have a significant impact on performance? Thank you.

How should I process the COLMAP dataset?

thanks for checking out spad!

using own dataset

if you wish to train spad on custom data, you can use the dataloader from zero123, which has the same coordinate system as spad.

for each item in your dataset, can easily compute plucker coordinates and epipolar mask using this function from spherical coordinates. these will be used during in the forward pass.

colmap dataset

you will have to preprocess colmap dataset by centering the points to origin and rescaling the scene to fixed bounds (say -1 to 1). also, you will have to convert them to blender/pytorch coordinate systems (see here and here).