Fig. 3: Recurrent motion integration in naturalistic scenarios.

a, Stage II-N refers to the outcomes from the Nth iteration in stage II, with visualizations illustrating neuron connectivity via heat maps. The neural connectivity is represented as a graph structure, and through graph bipartitioning¹⁵, the model can further achieve instance segmentation without any additional training. b, A scatter plot comparing human and model responses to the Sintel dataset in terms of u, v (pixels per frame). A red regression line, fitted to the data, indicates a strong linear relationship between human and model responses. The shadows around the red line represent the 95% confidence interval (CI) of the linear regression line. c, Qualitative comparison of ground truth, human and model responses on the MPI-Sintel²⁸. The larger the red circle at each location, whose size indicates the magnitude of positive RCI, the stronger the alignment with human responses over ground truth. At many points, our model demonstrates a better alignment with human perception than with ground truth. See Table 1 for detailed quantitative results. The model only uses the first-order channel in stage I. vs, versus.