Expanding the V1-MT model to the estimation of perceived fluid direction

Abstract

Humans can readily perceive the direction of liquid flow, yet computational modeling of this process remains challenging due to the complexity of non-rigid motion. Previous models based on neural activities in the primary visual cortex (V1) and the middle temporal area (MT) have been effective in explaining rigid motion perception. In this study, we extend the V1-MT model to address the perception of liquid flow direction. Participants observed video clips of liquid flow and reported the perceived direction, while the V1-MT model was used to predict these perceptions. The winner-take-all approach failed to accurately capture the observed perceptions. In contrast, a weighted mean of directional energies yielded strong predictions, highlighting that the human visual system spatially integrates directional energies from non-rigid motion components. These findings broaden the applicability of the V1-MT model to non-rigid motion and provide insights into how the visual system bridges the gap between computational models of rigid and non-rigid motion perception.

Keywords: Liquid flow direction; Motion; V1-MT model; Weighted average.

Fig. 1

(a) The processing pipeline of the V1-MT model used in the present study, illustrating the stages from spatiotemporal quadrature filtering to the computation of direction energy. (b) The distribution of direction energy across selective directions of MT neurons for different noise motion directions. (c) A box plot comparing the absolute errors in motion direction estimation between the winner-take-all and weighted averaging models.

Fig. 2

(a) Left: Snapshots of stimulus clips. Right: A snapshot of the experimental display showing a video frame of a stimulus clip accompanied by an arrow used by participants to report the perceived direction of liquid flow. (b) The red line represents the mean perceived direction for a stimulus clip across participants, while the gray thin arrows indicate individual perceived directions. The green dashed line and blue dotted line represent the directions inferred by the winner-take-all and weighted averaging models, respectively.

Fig. 3

(a) Direction energies summed across MT neurons and fitted with a double von Mises function. The green dashed line and blue dotted line represent the directions inferred by the winner-take-all and weighted averaging models, respectively. (b) Box plots of the absolute errors between participants’ perceived direction and the inferred directions from the winner-take-all (left) and weighted averaging (right) models.

Fig. 4

(a) Reported and inferred directions of liquid flow for each clip, plotted as a function of the start frame. Each clip consists of a 5-frame sequence of images. The values in the titles of the graphs indicate the absolute errors between the median reported direction and the inferred direction, as estimated by the weighted averaging (WA) and winner-take-all (WTA) models. (b) Box plots showing the absolute differences between reported and inferred directions for both the WA and WTA models across all clips.