Mid-Level Features Elicit Cognitive and Neural Representations of Object Size

Abstract

Most models of object recognition assume that we first recognize objects at the basic-level (e.g., as a “cup”), and then that the resulting object representations act as pointers that allow us to access knowledge about these objects, e.g., “can be held with one hand” or “is not alive” (e.g., DiCarlo & Cox, 2007). This dissertation advances an alternative view that recognition need not be a bottleneck for the visual system to infer certain kinds of knowledge about objects. In particular, I argue that cues to an object’s size in the real world may be read out from mid-level perceptual features (e.g., texture and form) that are processed prior to basic-level recognition. Chapter 1 serves as an existence proof that small objects (e.g., cups, books) and big objects (e.g., cars, couches) differ in mid-level perceptual features. Here, we use visual search performance as an index of perceptual similarity, and find that big and small objects differ in perceptual features even when we create unrecognizable versions of big and small objects that preserve texture and form information (“texforms”). In Chapter 2, we find that even though these texforms cannot be identified at the basic-level (e.g., as a “car”), they give rise to a Size-Stroop effect (Konkle & Oliva, 2012), suggesting that these mid-level features automatically trigger the processing of object size. Furthermore, we isolate perceived curvature as a feature that observers used to infer the real-world size of unrecognizable texforms. Finally, in Chapter 3, we find that unrecognizable texforms are sufficient to elicit the large-scale organization of object-selective cortex by object size (Konkle & Caramazza, 2013). Taken together, these results suggest that observers may use mid-level perceptual features to infer the size of an object in the real world, and that these same features can drive activity in regions that typically process recognizable images of big objects or small objects. The dissertation ends by exploring the developmental origins of object size representations.