The human brain assembles three-dimensional (3D) percepts from qualitatively different depth cues, yet the perceived 3D structure that the brain builds--a representation shared across cues--has remained difficult to measure directly. Here, we show that this cue-invariant 3D structure can be externalized as explicit 3D objects from human brain activity: fMRI responses are decoded into the latent features of a pretrained 3D point-cloud autoencoder, and a generator then maps these features back to a point cloud. A decoder trained exclusively on responses to 2D rendered objects passed three increasingly stringent tests: (i) it generalized to novel object categories; (ii) it generalized across depth cues to random dot stereograms (RDSs), which evoke 3D percepts through binocular disparity but share no pictorial shape information with the training images; and (iii) it tracked the 3D slant of contour-matched RDSs whose 2D outlines were held identical but whose disparity-defined slants varied, indicating that the reconstruction reflected depth-defined geometry rather than object category or image outline. Cross-cue generalization was strongest in higher visual areas, particularly along the dorsal stream. These results indicate that cross-cue generalization can serve as a criterion for externalizing perceived 3D structure and open a route toward reading out internal 3D representations that go beyond the momentary retinal input and could support predictions of how the world would appear under different viewpoints--a step toward externalizing the brains internal world model.
Aoki, S. C. et al. · CC-BY 4.0