Figure 3—figure supplement 2.

(A) Performance in predicting different generative parameters of point clouds in a synthetic dataset from either a raw or a size- and rotation-corrected (cell frame of reference, CFOR) embedding. As expected, CFOR normalization removes all information on cloud size ('scaling') and orientation ('rotation' 1–2). Interestingly, removing this information allows the regressor to perform far better when it comes to the shape parameters of the point cloud ('shape' 1–14). The 'random' parameter is a random Gaussian distribution and serves as a negative control. The regressor used is a Support Vector Regressor (SVR) with an RBF-kernel. (B) Evaluation of CBE compared to an alternative embedding strategy based on moments. Shown is how well the parameters used to synthetically generate point clouds can be predicted from embeddings of said clouds using different regression models (kNN: k-Nearest Neighbor regressor, l-SVR: linear Support Vector Regressor, r-SVR: RBF-kernel Support Vector Regressor). Black dots indicate results of 3-fold cross-validation, bars indicate the mean. Moments-based embedding is outperformed by CBE in all cases except with linear SVR, where the results are similar.