The concurrent encoding of viewpoint-invariant and viewpoint-dependent information in visual object recognition

ABSTRACT

A major theme of Glyn Humphreys’ career was object constancy, defined in his paper with Jane Riddoch (1984, p. 385) as “the ability to recognize that an object has the same structure across changes in its retinal project.” In this seminal neuropsychological work, they posit that there may be two routes to object constancy: one using local distinctive features and one based on global structure. Much of the work following Humphreys has focused on a similar dichotomy: whether recognition is viewpoint-invariant or viewpoint-dependent. This question has been debated at length, but has never been resolved in that, under different circumstances, both view-dependent and view-invariant recognition behaviour have been observed. We hypothesize that these inconsistent patterns can be accounted for by object representations that encode information supporting both patterns of recognition performance whereby the nature of the information expressed during recognition depends upon context. Our present results establish that viewpoint-invariant and viewpoint-dependent information is encoded concurrently, even if one information type is not immediately relevant to encoding context. In Experiment 1, participants concurrently, driven by the discrimination context, recruited viewpoint-invariant information for some objects and viewpoint-dependent information for others. In Experiments 2 and 3, participants learned the identities of novel objects that could be differentiated on the basis of view-invariant information and displayed viewpoint-invariant behaviour over changes in view; however, when recognizing these now-familiar objects in the context of new objects that were visually similar, participants shifted to displaying viewpoint-dependent behaviour. This viewpoint dependence was related to the particular views had been seen earlier in the experiment, even though at that time recognition was viewpoint invariant. Our results indicate that object representations are neither viewpoint-dependent nor viewpoint-invariant, but rather encode multiple kinds of information that are deployed in a flexible manner appropriate to context and task.

KEYWORDS:

• Object constancy
• view-invariance
• view-dependence
• object representation

Disclosure statement

No potential conflict of interest was reported by the authors.

ORCID

Michael J. Tarr http://orcid.org/0000-0003-4724-1744

Notes

1. Note that Just and Carpenter (1985) also proposed a model for when and why mental rotation is used in solving certain visual tasks, again involving both viewpoint-invariant and viewpoint-dependent information. However, their study also specifically relied on Shepard and Metzler’s (1971) 3D objects whereby the critical discrimination was of handedness (distinguishing left from right; Corballis & McLaren, 1984) and, as such, as discussed in Tarr and Pinker (1989), may not reflect the kinds of discriminations typically involved in everyday object recognition (unless one is an organic chemist). In addition, as pointed out by Takano (1989), Just and Carpenter’s (1985) account does not present any particular theory – in essence they provide a tautology in which viewpoint-invariant mechanisms are used when representations are viewpoint-invariant and viewpoint-dependent mechanisms are used when representations are viewpoint-dependent.

2. All of the shape contrasts beyond 1D for our stimuli were 2D, not 2D+. Tarr and Pinker (1990) referred to these contrasts as “2D+” because of their hypothesis that any dimensionality greater than 2D would also lead to viewpoint-dependent recognition.

3. This slightly off from upright orientation was used to ensure that the first training orientation did not bias participants to rely on viewpoint-dependent mechanisms more than they otherwise might. That is, shapes in the canonical 0° orientation have all parts aligned with an observer's gravitational and ground-plane axes. Thus, this orientation may be better recognized independent of training, see Tarr, Williams, and Pinker (1989).

4. If anything, such large shape differences among the unique central parts of the six target objects would seem to tilt the scales in favour of viewpoint-independence, the opposite of what was actually observed in the surprise phase of Experiment 3.

5. Indeed, interpreting view invariant patterns of recognition behaviour – essentially a null result in which there is no effect of the independent variable (viewpoint) – has been a recurring issue in the field (Tarr, Williams, et al., 1998).

6. There is, of course, an extensive body of neurophysiological research on the encoding of viewpoint in single neurons or populations of neurons (Logothetis, Pauls, Bülthoff, & Poggio, 1994; Ratan Murty & Arun, 2015). However, it is still unclear how the responses of individual neurons relate to behaviour (although see Perrett, Oram, & Ashbridge, 1998).