# Learning Continuous Image Representation with Local Implicit Image Function

**Authors:** Yinbo Chen, Sifei Liu, Xiaolong Wang  
**arXiv:** [2012.09161](https://arxiv.org/abs/2012.09161)  
**Date:** 2020-12-16  
**Summary page:** https://www.summarizepaper.com/en/paper/2012.09161v2/

## Key points

- Representing images in a continuous manner is challenging due to machines processing visual information using 2D arrays of pixels
- Recent advancements in 3D reconstruction with implicit neural representation have led to the development of Local Implicit Image Function (LIIF)
- LIIF involves using image coordinates and surrounding deep features to predict RGB values, allowing for arbitrary resolution presentation
- An encoder trained with LIIF representation through self-supervised tasks like super-resolution enhances continuous image representation
- LIIF can extrapolate learned representations to x30 higher resolution levels even without explicit training tasks
- LIIF bridges the gap between discrete and continuous representations in 2D, supporting learning tasks with size-varied image ground-truths
- Authors Yinbo Chen, Sifei Liu, and Xiaolong Wang's study on LIIF's efficacy has been recognized at CVPR 2021 (oral) and includes a project page for further exploration
- LIIF outperforms methods involving resizing ground-truths, showcasing its potential to revolutionize image representation and processing in computer vision applications

## Summary

In the realm of computer vision, representing images in a continuous manner poses a challenge due to machines storing and processing visual information in a discrete fashion using 2D arrays of pixels. Seeking to bridge this gap, recent advancements in 3D reconstruction with implicit neural representation have inspired the development of the Local Implicit Image Function (LIIF). This innovative approach involves taking an image coordinate and surrounding 2D deep features as inputs to predict the RGB value at that specific coordinate, allowing for arbitrary resolution presentation. To further enhance the continuous representation of images, an encoder is trained with LIIF representation through a self-supervised task involving super-resolution. The resulting learned continuous representation can be extrapolated to x30 higher resolution levels even when training tasks are not explicitly provided. Notably, LIIF serves as a pivotal link between discrete and continuous representations in 2D, enabling support for learning tasks with size-varied image ground-truths. The study conducted by authors Yinbo Chen, Sifei Liu, and Xiaolong Wang delves into the efficacy of LIIF representation in enhancing image processing tasks. Their work has been recognized at CVPR 2021 (oral) and includes a project page featuring videos and code for further exploration. By demonstrating superior performance compared to methods involving resizing ground-truths, LIIF showcases its potential to revolutionize how images are represented and processed in computer vision applications.