Staying in Shape: Learning Invariant Shape Representations using Contrastive Learning

Creating representations of shapes that are invari-ant to isometric or\nalmost-isometric transforma-tions has long been an area of interest in shape\nanal-ysis, since enforcing invariance allows the learningof more effective and\nrobust shape representations.Most existing invariant shape representations\narehandcrafted, and previous work on learning shaperepresentations do not focus\non producing invariantrepresentations. To solve the problem of\nlearningunsupervised invariant shape representations, weuse contrastive\nlearning, which produces discrimi-native representations through learning\ninvarianceto user-specified data augmentations. To producerepresentations that\nare specifically isometry andalmost-isometry invariant, we propose new\ndataaugmentations that randomly sample these transfor-mations. We show\nexperimentally that our methodoutperforms previous unsupervised learning\nap-proaches in both effectiveness and robustness.\n