Collator units: second-stage orientational filters

Abstract

Which parts of an image belong to the other parts? This has become known as the 'binding problem' and is almost as resistant to explanation now as it was over half a century ago, when the purely descriptive Gestalt rule of 'good continuation' and the like were formulated. The specific case of interest here is that of spatially sampled fragments of a continuous but partly occluded line. First-stage orientated visual filters with relatively small receptive fields (such as simple cells) make precise but essentially local measurements of the orientational properties of line fragments within their spatial domain. I suggest that these units act as 'tributary units', their outputs becoming the inputs to 'collator units', which are second-stage oriented filters whose function is to put together the local descriptions of the spatially distributed line fragments. The functional receptive field of the collator units would be the spatial sum of the receptive fields of the tributary units. If the tributary units all have the same orientation preference and their fields are axially aligned end to end, the receptive field of the corresponding collator unit would have the same orientation preference and width as its tributary units, but would be considerably longer. Psychophysical data are presented which are consistent with this model.