TL;DR
This paper introduces GT-mean loss, a new loss function that addresses brightness mismatch in low-light image enhancement, leading to improved model performance without significant computational overhead.
Contribution
The paper proposes the GT-mean loss, a simple and effective method to mitigate brightness mismatch in supervised low-light image enhancement models.
Findings
GT-mean loss improves performance across various datasets.
Incorporating GT-mean loss enhances existing LLIE methods.
The method introduces minimal additional computational costs.
Abstract
Low-light image enhancement (LLIE) aims to improve the visual quality of images captured under poor lighting conditions. In supervised LLIE research, there exists a significant yet often overlooked inconsistency between the overall brightness of an enhanced image and its ground truth counterpart, referred to as brightness mismatch in this study. Brightness mismatch negatively impact supervised LLIE models by misleading model training. However, this issue is largely neglected in current research. In this context, we propose the GT-mean loss, a simple yet effective loss function directly modeling the mean values of images from a probabilistic perspective. The GT-mean loss is flexible, as it extends existing supervised LLIE loss functions into the GT-mean form with minimal additional computational costs. Extensive experiments demonstrate that the incorporation of the GT-mean loss results…
Peer Reviews
Decision·ICLR 2025 Conference Withdrawn Submission
1. Proposes a straightforward loss function that effectively mitigates brightness mismatches in LLIE. 2. Demonstrates performance improvements across various LLIE models, supporting the generalizability of GT-Mean Loss. 3. The approach adds minimal computational overhead, making it easy to integrate into existing models without significant resource costs.
The authors claim that "brightness mismatch dominates the PSNR values," suggesting that brightness inconsistency heavily biases this commonly used metric, leading to inaccurate quality evaluations. However, this assertion may oversimplify the limitations of PSNR, as the issue presented seems to be a specific type of counterexample rather than an overarching flaw in PSNR itself. PSNR is limited primarily in its sensitivity to perceptual qualities rather than simply brightness mismatch, and as suc
1. The proposed loss leads to consistent performance for several backbones on three low-light image enhancement datasets. 2. The paper is well-motivated and elaborately develops a loss function for the observation.
1. In Fig.2, it seems more like a metric problem. If PSNR is not sensitive to this mismatch, why can the proposed loss function improve the PSNR performance? When there is a resolution difference between the original GT and the scaled image, how do we compute the PSNR value of the scaled one? Moreover, how is the brightness mismatch problem illustrated in Fig.2? 2. The paper couples the original loss and the introduced one using W. How do we confirm that these two loss functions have any relatio
1. The proposed method is well-motivated. The MSE/L1-based metrics, say PSNR or RMSE, do have a problem identifing noise from lightness bias in RGB color space. 2. The experiment results are reasonable.
1. Novelty. The GT-mean metric was first proposed in KinD[1], and has been adopted by many previous methods[2,3]. I don't think the effectiveness of GT-mean **metric**, can be considered as a contribution. 2. Formulation of global lightness. The author models the global brightness with normal distribution, which breaks in most real-world scenarios. [1] Zhang et. al, Kindling the Darkness: A Practical Low-light Image Enhancer, ACM MM 2019. [2] Wang et. al, Low-Light Image Enhancement with Norma
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
