Abstract
Objects can be altered by a wide variety of shape-transforming processes, such as bending, denting, twisting or stretching. Estimating what transformations have been applied to an object solely from its shape, is important for many other tasks (e.g., inferring material properties). Despite this, little is known about which cues observers use to infer shape transformations. Here, we sought to identify geometrical features that drive the identification of nonrigid deformations. We generated and rendered a large set of transformed unfamiliar 3D objects by shifting their mesh vertices around using smooth 3D deformation vector fields. Then, we compared observers’ inferences to predictions derived from 2D and 3D shape features. In the first experiment, we defined a transformation space spanned by weighted combinations of three transformation types (“twist”, “shear”, and “squeeze”). In each trial, observers viewed a test object generated at one of six positions in the transformation space. They adjusted the transformation applied to a different object, using the mouse to navigate within the triangular transformation space, until it matched the test. Observers showed remarkable invariance in their ability to infer transformations across objects. In the second experiment, we generated a more diverse set of stimuli by subjecting a larger variety of objects to 6 vector field transformations. For this set, there was more variance in performance with respect to transformation ground truth of the test object. We compared human responses to a computational model based on 2D and 3D shape features, finding that observers base their judgments on combinations of shape features that are reliable indicators of the transformation. These findings demonstrate that visual inferences of transformation processes rely on processes of perceptual organization and on hallmark transformation features, similar to visual processing in object perception.
Acknowledgement: This research was supported by an ERC Consolidator Award (ERC-2015-CoG-682859: “SHAPE”) and by the DFG (SFB-TRR-135: “Cardinal Mechanisms of Perception”)