LDINet: Latent Decomposition and Interpolation for Single Image FMO Deblatting

Haodong Fan, Dingyi Zhang, Yunlong Yu, Yingming Li

Primary Area: representation learning for computer vision, audio, language, and other modalities

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Keywords: fast moving object deblatting, image deblur, time super-resolution, latent decomposition and interpolation

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.

Abstract: The image of fast-moving objects usually contains a blur stripe indicating the blurred object that is mixed with backgrounds. To deblur the stripe and separate the object from the background in this single image, in this work we propose a novel LDINet that introduces an efficient decomposition-interpolation module (DIB) to generate the appearances and shapes of the objects. In particular, under the assumption that motion blur is an accumulation of the appearance of the object over exposure time, in the latent space the feature maps of the long blur is decomposed into several shorter blur parts. Specifically, the blurry input is first encoded into latent feature maps. Then the DIB module breaks down the feature maps into discrete time indexed parts corresponding to different small blurs and further interpolates the target latent frames in accordance with the provided time indices. In addition, the feature maps are categorized into the scalar-like and gradient-like classes which help the affine transformations effectively capture the motion of feature warping in the interpolation. Finally, the sharp and clear images are rendered with a decoder. Extensive experiments are conducted and has shown that the proposed LDINet achieves superior performances compared to the existing competing methods.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 5383