This image shows a research poster presented at the CVPR (Computer Vision and Pattern Recognition) conference held from June 17-21, 2024, in Seattle, WA. The poster details a study titled "Zero-Reference Low-Light Enhancement" by Wenjing Wang, Huan Yang, and Jianlong Yang. The focus of the research is to develop a method for enhancing low-light images using normal light images, eliminating the need for supervised learning.

The introduction section outlines the aim to improve robustness in terms of data usage, illumination-specific hyper-parameters, and unseen scenarios. The methodology involves designing an illumination-invariant prior to bridge the gap between normal and low-light images.

The method section describes the training framework, which involves predicting a physical quadruple prior and reconstructing it back to images. An included diagram demonstrates the process of training on normal light images and inferring on low-light images.

Examples of low-light enhancement effects are provided, showcasing the improvements made by the proposed method in various test images. Additionally, the setup includes a detailed breakdown of the different processing stages and their outcomes, illustrating the effectiveness of the zero-reference approach. 

The poster is part of a stand-alone display, with a small portion of the researcher's feet visible at the bottom, indicating they are at the venue ready to present their work.
Text transcribed from the image:
CVPR
JUNE 17-21, 2024
SEATTLE, WA

Zero-Reference Low-Light Enhancement
Wenjing Wang¹  Huan Yang²  Jianlong Yang²

Introduction

Zero-Reference Low-Light Enhancement: learn solely with normal light images, reducing the need for supervision

Our aim: improve the robustness to
- Data usage during training
- Illumination-specific hyper-parameters
- Unseen scenarios

Our methodology: design an illumination-invariant prior that serves as a bridge between normal light and low-light images

Training on normal light images
Normal Light Input
Physical Quadruple Prior
Illumination Invariant Features
Prior-to-Img Framework
Output
Reconstruction Loss

Inference on low-light images
Low-Light Input
Physical Quadruple Prior
Illumination Invariant Features
Prior-to-Img Framework
Output

Method

Training framework
- Predict a physical quadruple prior
- Reconstruct the prior back to image
Layers
Gaussian Convolve
SD Encoder

Solution of detail degradation: bypass
SD Decoded

Low-light enhancement effects for differen
Test Image
w/o H
w/o C