BVI-Lowlight: Fully registered datasets for low-light image and video enhancement

Low-light images and video footage often exhibit issues due to the interplay of various parameters such as aperture, shutter speed, and ISO settings. These interactions can lead to distortions, especially in extreme lighting conditions. This distortion is primarily caused by the inverse relationship between decreasing light intensity and increasing photon noise, which gets amplified with higher sensor gain. Additionally, secondary characteristics like white balance and color effects can also be adversely affected and may require post-processing correction. These distortions not only impact the perceived quality of the images but also pose significant challenges for machine learning tasks, including classification and object detection. This is particularly evident when considering the susceptibility of deep learning networks to adversarial examples.The BVI-Lowlight datasets offer fully registered low-light content alongside their corresponding clean and normal light condition. This dataset includes both images and videos, enabling the use of supervised learning approaches and performance evaluation through objective metrics such as PSNR and SSIM.Two datasets are available:BVI-RLV: Fully Registered Low-Light Videos (BVI-Lowlight-videos):In this video pair dataset, we recorded low-light videos at both 10% and 20% of normal lighting levels (100%), indicated by the Zero 88 FLX S24 light controller. We provide these videos in full HD resolution. There are total 40 scenes, including 6 scenes of static background. More detail at: https://arxiv.org/abs/2407.03535BVI-Lowlight-Images:The description can be found on https://github.com/malalejandra/bvi-lowlightBenchmarks for Low-Light Video Enhancement:PCDUNet: https://github.com/lrr-rachel/PCDUNetSTA-SUNet: https://github.com/lrr-rachel/STA-SUNetBVI-CDM: https://github.com/lrr-rachel/BVI-CDMBVI-Mamba: https://github.com/russellllaputa/BVI-MambaPlease cite R. Lin, N. Anantrasirichai, G. Huang, J. Lin, Q. Sun, A. Malyugina, and D.R. Bull. BVI-RLV: A fully registered dataset and benchmarks for low-light video enhancement. arXiv preprint arXiv:2407.03535, 2024.

低照度图像及视频素材常因光圈、快门速度及ISO设置等参数的相互作用而出现各类问题。此类交互作用在极端光照条件下尤为显著，可导致图像畸变。此类畸变主要由随光照强度减弱而光子噪声增强的反向关系引起，且随着传感器增益的提升而加剧。此外，如白平衡和色彩效果等次生特性亦可能遭受不利影响，并可能需要后期处理进行校正。这些畸变不仅影响图像的感知质量，亦对机器学习任务构成重大挑战，包括分类和目标检测等。尤其是在考虑深度学习网络对对抗样本的敏感性时，此问题尤为突出。BVI-Lowlight数据集提供与其相对应的清洁及正常光照条件下完全配准的低照度内容，包括图像和视频，从而支持监督学习方法的运用，并通过如峰值信噪比（PSNR）和结构相似性（SSIM）等客观指标进行性能评估。本数据集包括两个子集：BVI-RLV：完全配准的低照度视频（BVI-Lowlight-videos）：在此视频对数据集中，我们记录了在正常光照水平10%和20%（即100%）下的低照度视频，由Zero 88 FLX S24灯光控制器指示。我们提供全高清分辨率的视频。总共有40个场景，包括6个静态背景场景。更多详情请见：https://arxiv.org/abs/2407.03535。BVI-Lowlight-Images：描述请见https://github.com/malalejandra/bvi-lowlight。低照度视频增强基准：PCDUNet：https://github.com/lrr-rachel/PCDUNetSTA-SUNet：https://github.com/lrr-rachel/STA-SUNetBVI-CDM：https://github.com/lrr-rachel/BVI-CDMBVI-Mamba：https://github.com/russellllaputa/BVI-Mamba请引用R. Lin, N. Anantrasirichai, G. Huang, J. Lin, Q. Sun, A. Malyugina, and D.R. Bull. BVI-RLV：低照度视频增强的完全配准数据集和基准。arXiv预印本arXiv:2407.03535，2024。