Fig. 6: The proposed events-to-frame alignment pipeline.

a The sparse events E are stacked into a 2-D frame E(x, y), and b denoised to \({E}^{{\prime} }(x,y)\), which is used for contour matching with the GF image. For the GF images, c several images are averaged together to remove the sand grain features, and a filter is used to enhance the contour. After d contour matching with \({E}^{{\prime} }(x,y)\), the centroid displacements Δx = + 5, Δy = − 8 and the scale factor s = 1.1 are retrieved, and applied to (e) spatially transform the (raw) real event data. To refine these alignment parameters, synthetic events are generated with an event emulator, and they serve as the GT to tune the parameters using the K-Nearest Neighbors (KNN) algorithm. The values in (f) refinement curves indicate the impact on the accuracy when the shifts, \(\Delta {x}^{{\prime} }=+6\), \(\Delta {y}^{{\prime} }=+1\), \(\Delta {s}^{{\prime} }=0\), are applied. All of these images have dimensions of 1160 × 680 pixels and a size of 1.1 μm. The scale bars, shown as white lines in the frames, represent 200 μm.