Learning to Orient Surfaces by Self-supervised Spherical CNNs

ConDor: Self-Supervised Canonicalization of 3D Pose for Partial Shapes

“…Most work on canonicalization evaluates performance indirectly on downstream tasks such as segmentation or registration [42,45]. This makes it hard to disentangle canon-Table 1.…”

“…This makes it hard to disentangle canon-Table 1. Full shape canonicalization compared to a PCA baseline, Canonical Capsules (CaCa) [45] and Compass [42], and full (F) and full+partial (F+P) versions of our method. We outperform methods on most categories and metrics.…”

“…We choose PCA as a baseline -for each shape we compute the top-3 principal components and use this as an equivariant frame for alignment across instances. We compare with two methods for rotation canonicalization: Canonical Capsules (CaCa) [45] and Compass [42]. Results for full shape canonicalization are shown in Table 1.…”

“…The problem is further exacerbated by the difficulty of canonicalizing partial shape observations (e.g., from depth maps [35]), or symmetric objects that require an understanding of inter-instance part relationships. Recent work addresses these limitations using weakly-supervised [14,37] or self-supervised learn-ing [28,42,45], but cannot handle partial 3D shapes, or is limited to canonicalizing only orientation.…”

“…We therefore propose four new metrics that are designed to evaluate the consistency of instance-and category-level canonicalization, as well as consistency with manually pre-canonicalized datasets. We extensively evaluate the performance of our method using these metrics by comparing with baselines and other methods [42,45]. Quantitative and qualitative results on common shape categories show that we outperform existing methods and produce consistent pose canonicalizations for both full and partial 3D point clouds.…”