Build my own dataset for training

[zrw97402]

Thanks for your excellent work first.
I'm working on my own dataset for free-viewpoint rendering and interested in your work. Since rgb, pose and intrinsic datas are prepared, how can I get the bbox data like the style in your datasets?

[MultiPath]

The bounding box is usually roughly estimated. For example, if you have the object mask, it is possible to run algorithm like visualhull to get the rough geometry. You can also assign a unit bounding box if you know the object's rough size.

[zrw97402]

Thanks for your answer. My target dataset is the open scenes such as a basketball game without a fixed boundary. I wonder if these can be constrainted into such a bbox since some values are unknown and uncertain.

[MultiPath]

It would be quite challenging to model unbounded scenes with the current version of NSVF. Since we are basically based on voxels and will prune them during training. So it currently only works in Euclidean space. For large unbounded scenes, it would very inefficient to work in such scenes. I would be interested in how it will look like.

[zrw97402]

I've got your words. Thank you.

[dwangxxx]

It would be quite challenging to model unbounded scenes with the current version of NSVF. Since we are basically based on voxels and will prune them during training. So it currently only works in Euclidean space. For large unbounded scenes, it would very inefficient to work in such scenes. I would be interested in how it will look like.

Hello, I've read your paper recently! It's an excellent work. But I have some questions about the details. In the training stage, the voxels are pruned to get more accurate voxels. I want to know if the pruned voxel are applied int the inference stage? Look forward your apply! Thank you!