Discussion on Visual feature aggregation

[vidit98]

Thanks again for releasing the code. I was going through extract_features.py and found that to calculate the feature for a voxel, visibility of the voxel is not considered from a given camera.

Intuitively, I thought that features would be averaged only from rendered views from which a particular voxel is available. Do have any explanation why does visibility does not matter? Did you guys tried something like this before finalizing the current method?

[YuDeng]

This is a good question. We haven't tried taking visibility into consideration.

In our case, a simple average over all views works pretty well. We conjecture that when averaging, the visible features corresponding to a voxel tend to dominate, given that objects are usually somewhat convex. Of course, we believe that incorporating visibility could yield even better results, and it is worth trying.