DeLiEve-Net: Deblurring Low-Light Images With Light Streaks and Local Events

Chu Zhou, Minggui Teng, Jin Han, Chao Xu, Boxin Shi; Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops, 2021, pp. 1155-1164

Abstract


Modern blind deblurring methods usually show degenerate performance when handling images captured in low-light conditions because these images often contain saturated regions of light sources, and the image contents and details in dark regions are poorly visible. In contrast, event cameras can faithfully record the positions and polarities of intensity changes with a very high dynamic range and low latency, which suffer less in the dark than conventional cameras. However, existing event-based deblurring methods require guidance from global events with the same spatial resolution as the blurry image (typically 346 * 260 pixels), which significantly limits the spatial resolution of images they can process. In this paper, we address this problem in a two-stage way by proposing a neural network named DeLiEve-Net, which learns to Deblur low-Light images with light streaks and local Events. An RGB-DAVIS hybrid camera system is built to validate that our method can deblur high-resolution RGB images with events in low-light conditions.

Related Material


[pdf] [supp]
[bibtex]
@InProceedings{Zhou_2021_ICCV, author = {Zhou, Chu and Teng, Minggui and Han, Jin and Xu, Chao and Shi, Boxin}, title = {DeLiEve-Net: Deblurring Low-Light Images With Light Streaks and Local Events}, booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops}, month = {October}, year = {2021}, pages = {1155-1164} }