Abstract
A number of models have been proposed over the years that are able to estimate the speed of a moving image feature such as an edge but it is not obvious how these models should be assessed in terms of their performance. Over what range of speeds should a model's estimates of image velocity be veridical in order for it to be classed as effective? There is currently a lack of data that can directly inform us as to what the function looks like that links human estimates of speed (v′) to actual speed (v), i.e., v′ = f(v), f = ? On a plot of v′ versus v, it is difficult to establish the absolute location of the function but we will show that there already exists a range of psychophysical data which constrain the form it can take. For example, the U-shaped, speed discrimination (Weber fraction) curves obtained by a number of researchers (e.g., McKee, Vis Res., 1981; De Bruyn & Orban, Vis Res.1988) suggest that the v′= f(v) function for moving edges is s-shaped with the maximum slope occurring at intermediate speeds (approx 4 – 16 deg/s). We have discovered that this s-shape is also predicted by models of speed estimation that feature speed-tuned Middle Temporal (MT) neurons and which incorporate a weighted vector average (centroid) stage (e.g., Perrone & Krauzlis, VSS, 2009). Because the range of speed tunings in MT is naturally constrained at both the high and low speed ends, the centroid estimate of the MT activity distribution is biased as a result of ‘truncation effects’ caused by these lower and upper bounds; speed estimates in the model are overestimated at slow input speeds and underestimated at high input speeds producing an s-shaped, v′= f(v) function similar to that predicted by the speed discrimination data.
JP & RK supported by a Royal Society of New Zealand Marsden Fund grant.