Computing albedo, illumination, and optical flow with spacetime-adaptive bandpass filtering
Steven R. Reuman · 1988
The Intensity Dependent Spread (IDS) operation (Cornsweet, 1985) was first introduced as a model of retinal spatial summation and human brightness perception. IDS is of interest in machine vision because it is a spatial bandpass filter which adapts locally to photon noise. IDS has recently been extended to spacetime as the Constant-Velocity IDS (CVIDS) operation. CVIDS performs spacetime adaptive bandpass filtering while remaining consistent with human spatial and temporal psychophysical data. This thesis applies CVIDS to machine vision by estimating albedo, illumination, and speed from the CVIDS output to Mondrian inputs. The following is shown: (1) Albedo, illumination, and speed can be estimated from the CVIDS response to one-dimensional step edges, assuming that albedo changes are abrupt relative to illumination changes. Simulations suggest that the statistics of the estimates are stationary, despite nonstationary input photon noise. (2) Two-dimensional optical flow can be completely constrained by applying two or more DOG filters with different sigma's to an image sequence, if at least two of the bandpass outputs have linearly independent gradients and obey the optical flow constraint equation at each point. IDS and CVIDS can be used the same way because they are bandpass filters, but they can also constrain Mondrian edge motion twice by separably encoding two frequency components in a single output image sequence. (3) IDS output to a two-greylevel input is parametrically equivalent to a log transform followed by DOG on the same input. IDS output is necessarily different from log-DOG output for more complex inputs because of IDS local adaptiveness. (4) Three variations on CVIDS are developed. An adaptive bandpass operation for input functions of time alone (f(t)) is derived from CVIDS. A spacetime bandpass operation exhibiting all CVIDS behavior except its adaptivity is also derived. Finally, CVIDS is extended to a vector operation allowing computation of speed, albedo, and illumination within one algorithm. (5) A new method of extending Land's Retinex theory to two dimensions is derived using line integration. Similarities between the method and CVIDS behavior are discussed.