Box center: (0.250, 0.250, 0.000)
Box side lengths: (2.005, 1.503, 1.503)
The benefits of using view dependence is that the additional information can improve model quality and also get more specular information accross. However if our training images are more biased towards one view we can have irregularities which can make the model overfit. Hence a balanced data sampling is necessary. For the given datasets and image size its hard to make out a difference, but theoretically these points will hold.
I only changed the chunk size to account for GPU memory limits.
