BC04 Scooping, Test Overfitting
With single and multi-sphere cases.
Created on July 24|Last edited on July 24
Comment
(07/24/2022) Test overfitting.
I did this because according to this (results with multi-sphere):
Some of the Direct Vector (MSE) methods seemed to have exploding training MSEs.
Well the results do seem to suggest that the methods can overfit, though some of the training curves seem to shoot upwards. I'm not sure why, though, and the effect seems more pronounced for the "ground truth" point clouds which for whatever reason seem much harder than the "observable" point cloud for the Direct Vector MSE method. For ToolFlowNet, it can learn from both variants (good...). The g.t. v02 seems worse than g.t. v01 which might be a way to consistently show that with more points allocated to the balls, the worse performance seems to be.
Results
Run set
6
Add a comment