Intrinsic viewpoint invariance of 3-D objects

B. Tjan, G. E. Legge

Research output: Contribution to journalReview articlepeer-review


Purpose. Controversy about the viewpoint-dependent nature of human object representation is hard to resolve because human performance depends on both perceptual representations and information constraints of the recognition task. Intrinsic properties of the objects can lead to viewpoint-dependent performance, even if the perceptual representation is viewpoint invariant (Liu, Z. Spatial Vision, 1996). The intrinsic viewpoint invariance of a task can be measured by an ideal observer and depends on two factors: Intra-object viewpoint invariance (a view with look-a-like neighbors) is a property of a single object, assessed by an ideal observer's accuracy in judging which of two noisy views of a single object is identical to a noisy target view of that object (view-discrimination task). Inter-object viewpoint invariance (a diagnostic view) is a property of a set of objects, assessed by an ideal observer's performance in judging which of two noisy images is the same object as a noisy target (object-discrimination task). Methods. Computer-simulated ideal observers for these two tasks had knowledge of the 3-D object models and range of possible viewing angles. Objects of simple geometric forms were used, some with known aspect graphs (Tarr, M., & Kriegman, D. Annual Meeting of the Psychonomic Society, 1992). For both tasks, we plot the accuracy of the ideal observer as a function of target orientation. For a given target orientation, low accuracy in the view-discrimination task, or high accuracy in the object-discrimination task, is indicative of high intra-object or inter-object viewpoint invariance, respectively. Results. We found that high-accuracy object discrimination does not always occur at views with high intra-object invariance. Changing distractor objects affects locations of inter-object invariance. Furthermore, orientations that lack intra-object invariance do not always coincide with visual "events" as depicted on aspect graphs. Conclusions. A "canonical view" may either mean a view that is most similar to other views of an object (intra-object invariance), or a diagnostic view that distinguish one object from another (inter-object invariance). As we have shown, these concepts are not equivalent. Thus, analyses based on the properties of individual objects are inadequate for describing the informational constraints on object recognition.

Original languageEnglish (US)
Pages (from-to)S1125
JournalInvestigative Ophthalmology and Visual Science
Issue number3
StatePublished - Feb 15 1996

Fingerprint Dive into the research topics of 'Intrinsic viewpoint invariance of 3-D objects'. Together they form a unique fingerprint.

Cite this