strictly as theories of high-level visual shape recognition. In particular, the 3D model does not have an intrinsically
visual character—it is no longer retinotopic,for example;nor,as Marr stresses, is itconfined totheobserver's pointof
view.^180
It is important to see how abstract the hypothesized level of SpS is. Although SpS is geometric in character, it is not
“imagistic”in the sense of the“percepts”of section 10.5: it is not to be thought of as encoding“pictures”or“statues
in the head.”An image is restricted to a particular point of view, whereas SpS is not. An image is restricted to a
particular instance of a category, whereas SpS is not (recall Berkeley's objection to images as the vehicle of thought:
howcanan image ofa particular trianglestand for allpossible triangles?). An image cannotinclude theunseenparts of
an object—its back and inside, and the parts of it occluded fro mthe observer's view by other objects—whereas SpS
does. An image is restricted tothevisual modality, whereas SpScan equallywellencode informationreceivedhaptically
or through proprioception. Nevertheless, even though SpSs are not themselves imagistic, it makes sense to think of
them as encoding image-schemas: abstract structures from which a variety of images can be generated and to whicha
variety of percepts can be compared.
The work of understanding the conceptualized world is divided between CS and SpS (and probably other levels).
Judgments and inferences having to do with predicate–argument relations, category membership, the type–token
distinction, quantification, and so forth can be formulated only in terms of CS. Judgmentsand inferences havingto do
with exact shapes, locations, and forces can be formulated only in terms of SpS. On the other hand, there is overlap
between the two levels, in that the notions of physicalobject, part–whole relationships, locations, force, and causation
have reflexes in both systems. It is these shared components that enable the two systems to communicate with each
otherthroughan interfaceoftheusualsort.Thisdivisionof labor isthus a moreabstractversionofPaivio's(1971) old
dual-coding hypothesis: CS is sort of
LEXICAL SEMANTICS 347
(^180) Some colleagues have objected to Marr's characterizing the 3D sketch as“object-centered,”arguing that objects are always seen fro mso me point of view or other—at the
very least theobserver's. However, I interpret“object-centered”as implyingonlythattheencoding of theobject is independent of pointof view. This neutralitypermits the
appearanceoftheobjectto be computedas necessary tofit theobjectintothevisual scene as a whole,viewed fro many arbitrary vantagepoint.Marr, who is notconcerned
withspatial layout, but onlywith identifyingtheobject, does not deal withthis further step of reinjecting the objectintothe scene. But I see such a step as altogetherwithin
the spirit of his approach.I recognize that Marr's 3D model is now out of fashion in the vision community. Still, the criteria for SpS must be satisfied bysomecognitive
structure or combinationof structures. Readers more conversantthan I in thevision literature should feelfreeto substitute their favoritetheoryof high-levelvision here, as
long as it satisfies the criteria we need to provide a“perceptual front end”for language.