Fulvio Domini

Associate Professor
Office Location: 
Metcalf 237
Research Focus: 
Computational vision

Fulvio Domini is a vision scientist who studies how the human visual system extracts 3D information from retinal images. He came to Brown in 1999 after completing his Ph.D. in experimental Psychology at the University of Trieste, Italy. He has taught courses on computational vision, perception, psychophysics and human factors.

Research Interests:

Human survival capabilities depend on a fundamental skill of the visual system: interpreting retinal images in order to extract knowledge about the 3D structure of the environment that surrounds us. Understanding how this process is carried out by the perceptual system has always been the main focus of Fulvio Domini research. In the last twenty years the vision community has approached this problem by making, implicitly or explicitly, three fundamental assumptions. First, since a mathematical analysis of different properties of retinal images, also called signals, indicates that these properties are sufficient for deriving the veridical structure of the objects that project on the retina there must be independent modules that perform such analysis. Second, these modules operate in a fairly instantaneous and spatially localized fashion. Third, the outputs of independent modules are combined at the very last stage of visual processing in order to provide a unique and veridical 3D representation of projected objects. In several published journal articles Fulvio Domini has provided empirical results, accounted for by alternative theoretical frameworks that challenge these assumptions. 

Fulvio Domini started his research career by investigating the first hypothesis. Specifically, he became intrigued with the problem of how the visual system extracts shape from dynamic properties of retinal images. It is well known that moving features on a flat image can induce a vivid impression of a three-dimensional structure. For many decades, theoretical and empirical investigations into this ability have interpreted these percepts as faithful copies of the projected 3D structures. In a series of investigations Fulvio Domini has provided empirical evidence that proves this interpretation non-biologically plausible. Moreover, he has developed a new theory of human processing of structure from motion that can predict previously reported findings and these new results. The most important and puzzling result that can be predicted by this new theory is that human observers judge local 3D properties of a smooth surface in a way that is inconsistent with how the surface is perceived as a whole. In other words he has provided empirical results that resemble the "ever ascending" staircase of the Dutch painter M.C. Esher. While this result is perfectly predicted on a local level by the model he developed, it seems to clash with the phenomenic experience of a perceived smooth global surface. 

This perceptual paradox made Fulvio Domini revisit again the second commonly held assumption described above and he concluded that the visual system does not analyze the image in a strictly local fashion. Similar considerations made him think that information for 3D shape is not processed instantaneously. He therefore received a NSF grant that allowed him to investigate processes of spatial or temporal integration for the perception of 3D shape. In a series of studies on this problem he discovered three important sets of findings. First, motion information for the perception of 3D structure is not processed instantaneously, but effects of temporal integration can be found for a temporal window that can span up to one second. Second, a similar process of temporal integration exists when binocular information is followed or preceded in time by motion information. Third, global structure from motion can be predicted by considering a process of spatial integration acting on local and non-veridical measurements of surface shape.

The results of Fulvio Domini's studies on structure-from-motion made him think again about the above-mentioned third hypothesis, that postulates a modular organization of 3D shape perception processing. The fact that structure-from-motion is, in general, non-veridical seems at odds with the idea of having a separate SFM module. What would the role of that module be? How can a non-veridical output be combined with the outputs of other modules? Moreover, he realized that investigations on depth-cue integration have been plagued with contradictory findings, where the reports consistent with the predictions of an optimal combination of independent depth estimates clashed with reports of interactions among depth modules. He is currently developing a new conceptual framework in which such contradictions may be resolved. According to this proposal, the visual system takes advantage of the natural co-variation that exists among image signals. Fulvio Domini calls these co-variations intrinsic constraints because they represent relationships among image measurements. These co-variations identify a lower dimensional manifold in the multi-dimensional space of signals. Domini's hypothesis is that the visual system derives 3D properties from measurements on this lower dimensional manifold. He is currently investigating this problem with the additional support of a grant that he has recently received by the National Science Foundation