Adaptive Motion Pooling and Diffusion for Optical Flow Computation

We study the impact of local context of an image (contrast and 2D structure) on spatial motion integration by MT neurons. To do so, we revisited the seminal work by Heeger and Simoncelli (HS) [4] using spatio-temporal filters to estimate optical flow from V1-MT feedforward interactions. However, the HS model has difficulties to deal with several problems encountered in real scenes (e.g., blank wall problem and motion discontinuities). Here, we propose to extend the HS model with adaptive processing by focussing on the role of local context indicative of the local velocity estimates reliability. We set a network structure representative of V1, V2 and MT areas of the motion stream. We incorporate three functional principles observed in primate visual system: contrast adaptation [3], adaptive afferent pooling [2] and MT diffusion that are adaptive dependent upon the 2D image structure (Adaptive Motion Pooling and Diffusion, AMPD). We evaluated both HS and AMPD models performance on Middlebury optical flow estimation dataset [1]. Our results show that the AMPD model performs better than the HS model and its overall performance is comparable with many modern computer vision methods. The AMPD model could be further improved by integrating feedback to better recover true velocities around motion boundaries. We propose that such adaptive model can serve as a ground for future research in biologically-inspired computer vision.