Ryne Choi^1,2 (), Jacob Feldman^1,2, Manish Singh^1,2; ¹Rutgers University - New Brunswick, ²Rutgers University, Center for Cognitive Science

In previous work, we demonstrated a violation of the rigidity assumption in Structure-From-Motion (SFM): a rigidly rotating plane, with one concave and one convex part protruding from each of its vertical halves, is perceived as a surface with two convex parts, moving non-rigidly (VSS, 2023, 2024). Under orthographic projection, observers predominantly perceived the non-rigid interpretation, while under perspective projection, non-rigid percepts were less frequent but still significantly above zero. We view these results as illustrating a competition between priors for rigidity and convexity, where convexity typically “wins” even in the presence of perspective cues — leading to non-rigid percepts despite all stimuli being rendered as rigid. We developed a model to construct a 3D shape re-interpretation where both parts are convex. The model “flips” points on the concavity across a mirror, along the ray from each point to the observer, thereby maintaining projective consistency, and transforming it into a convexity. For orthographic and perspective stimuli, the model differs in two important ways to maintain projective consistency: (1) the rays to the observer, reflecting differences in the image formation process; and (2) mirror location, as perspective provides cues to preserve the rigidity of the square plane, such as foreshortening. The model correctly predicts the non-rigid motion percepts that arise when observers perceptually reinterpret concave parts as convex. It also captures the qualitative change in non-rigid percepts between orthographic and perspective projections that we observed in our previous work. On the other hand, this pattern of results could not be explained by a simple motion-based heuristic that uses relative motion to determine the depth of each point. The success of the shape-based model underscores the significant role of the convexity bias in interpreting SFM: a shape prior is capable of predicting non-rigid motion percepts on a rigidly rendered SFM stimulus.