Learning Generalizable Part-based Feature Representation for 3D Point Clouds (Supplementary Material)

As defined in Eqn.(8), the overall training loss of PDG consists of the shape loss L As shown in Fig. S-1 (a), the performance of PDG is not sensitive to λ And in Fig. S-1 (b), the standard deviation of nine classification results is 0.66, demonstrating that our PDG is also not sensitive to λ We can observe that PointNet achieves relatively high test error and MetaSets (PointNet) could not sufficiently decrease the test errors. Compared with them, the test error of our proposed PDG (PointNet) decreases to a low value. All experiments are conducted on a single Tesla V100 GPU with batch-size 32. We can observe that PDG (PointNet) achieves the best balance on performance and computation cost. Compared with MetaSets (PointNet), our proposed PDG (PointNet) improves its classification accuracy by 7.3% while taking only respectively 3.6% and 86.7% time and space cost of MetaSets (PointNet).