All-in-Focus Imaging from Event Focal Stack

Defocused Image Event Focal Stack Neighbouring Events Image Focal Stack Predicted Weights All-in-focus Image (a) Defocused Image (b) Ground Truth (c) DRBNet (d) Ours Patches Refocusing Timestamp Selection Reconstruction Merging 𝑡𝑡 0 𝑡𝑡 4 𝑡𝑡 1 𝑡𝑡 0 𝑡𝑡 4 𝑡𝑡 1 + Reconstructed Sharpness 𝑡𝑡 𝑖𝑖 for patch 𝑖𝑖 Golden Search Iteration Figure 1. Left: We propose the event focal stack composed of event streams, which can be used to reconstruct an image focal stack and predict the merging weights for all-in-focus image recovery. Our pipeline consists of three steps: selecting the refocusing timestamps, reconstructing the corresponding image focal stack, and merging the stack into an all-in-focus image with weights predicted from the images and neighbouring events. Right: Given a defocused image (a) and the corresponding event focal stack, our method recovers an all-in-focus image (d) with closer clarity to the ground truth (b) than DRBNet [31] (c).