Proceedings of the Annual Meeting of the Cognitive Science Society

The human visual system robustly reconstructs 3D scenes from fleeting glances and sparse 2D inputs, effectively solving an ill-posed problem. In contrast, machine vision systems like 3D Gaussian Splatting often fail under such conditions, suffering from severe overfitting. We hypothesize this gap stems from the absence of biologically plausible inductive biases: specifically, Marr's 2.5D sketch and Gestalt organizational principles. To test this, we operationalized these theories within a computational framework, incorporating surface orientation cues (Marr) and the Gestalt principle of Pr–âgnanz to enforce structural simplicity and suppress redundancy. Experiments demonstrate that these cognitive constraints are computationally necessary to resolve geometric ambiguities in sparse-view environments. Crucially, our biologically constrained model yields representations judged by humans as significantly more realistic and structurally coherent than unconstrained baselines. These findings provide compelling computational evidence supporting the functional necessity of intermediate geometric representations and global organization principles in facilitating robust human 3D perception.