The monkey inferotemporal cortex (IT) is the association cortex implicated in object perception and recognition, but little is known how neurons in this area represent object images. Intrinsic signal imaging revealed that visual images of objects activated columns in a distributed manner in IT. When a part of these object were removed (stimulus simplification), the simplified stimuli activated only a subset of columns elicited by the originals. This result and subsequent extracellular recordings of neuronal activity from these columns suggested that objects were represented by the combination of columns, each of which represents a visual feature of objects (the feature-based representation). These features can be local features such as protrusions, curvature, and rectangular shapes appeared in part of object images. To find activity related to spatial relationship among the local features, we segmented an object image into parts where local features are accommodated and investigated spatial relationship among these parts instead of that among particular local features. Intrinsic signal imaging revealed the spots that were activated by the object but not by either of the individual parts. Extracellular recording of neuronal activity from these spots revealed that neurons were not sensitive to shapes of the individual parts, but to spatial arrangements of two parts. This result indicates that some columns represent local features but other represent global arrangements of such local features. Unique representation of an object may be achieved by combining both types of feature columns together. We have shown that representation of object images in monkey IT cortex is based on population coding with visual features of object images. One advantage of this feature-based representation is that a variety of possible combinations of visual features made it possible to represent many objects differently in IT cortex. If majority of visual features representing an object do not largely change in a certain range of different viewing angles, population coding would also help categorization of images of the object viewed from different angles. We examined this possibility by recording population activity elicited by various object images with a dense multiple-electrode array, and found that not single site activity but the population activity indeed helps to represent object images view invariantly for a certain range.