Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

BVI-Lowlight: Fully Registered Benchmark Dataset for Low-Light Video Enhancement

Nantheera Anantrasirichai, Ruirui Lin, Alexandra Malyugina, David Bull

TL;DR

The paper tackles the scarcity of aligned ground-truth data for low-light video enhancement by introducing BVI-Lowlight, a fully registered RGB video dataset with 40 dynamic scenes captured under two low-light levels ($10\%$ and $20\%$) and a normal-light ground truth ($100\%$). It provides controlled, repeatable motion via a motorized dolly and ground-truth alignment through histogram matching, enabling pixel-wise registration and robust full-reference evaluation. The authors benchmark multiple architectures, including UNet, CGAN, SwinIR, and their PCDUNet, demonstrating clear gains when training on registered data and under varying illumination conditions. They show the dataset’s usefulness for supervised learning and cross-dataset evaluation, and compare against existing datasets to establish its representative quality and practical impact for low-light video enhancement research.

Abstract

Low-light videos often exhibit spatiotemporal incoherent noise, leading to poor visibility and compromised performance across various computer vision applications. One significant challenge in enhancing such content using modern technologies is the scarcity of training data. This paper introduces a novel low-light video dataset, consisting of 40 scenes captured in various motion scenarios under two distinct low-lighting conditions, incorporating genuine noise and temporal artifacts. We provide fully registered ground truth data captured in normal light using a programmable motorized dolly, and subsequently, refine them via image-based post-processing to ensure the pixel-wise alignment of frames in different light levels. This paper also presents an exhaustive analysis of the low-light dataset, and demonstrates the extensive and representative nature of our dataset in the context of supervised learning. Our experimental results demonstrate the significance of fully registered video pairs in the development of low-light video enhancement methods and the need for comprehensive evaluation. Our dataset is available at DOI:10.21227/mzny-8c77.

BVI-Lowlight: Fully Registered Benchmark Dataset for Low-Light Video Enhancement

TL;DR

The paper tackles the scarcity of aligned ground-truth data for low-light video enhancement by introducing BVI-Lowlight, a fully registered RGB video dataset with 40 dynamic scenes captured under two low-light levels ( and ) and a normal-light ground truth (). It provides controlled, repeatable motion via a motorized dolly and ground-truth alignment through histogram matching, enabling pixel-wise registration and robust full-reference evaluation. The authors benchmark multiple architectures, including UNet, CGAN, SwinIR, and their PCDUNet, demonstrating clear gains when training on registered data and under varying illumination conditions. They show the dataset’s usefulness for supervised learning and cross-dataset evaluation, and compare against existing datasets to establish its representative quality and practical impact for low-light video enhancement research.

Abstract

Low-light videos often exhibit spatiotemporal incoherent noise, leading to poor visibility and compromised performance across various computer vision applications. One significant challenge in enhancing such content using modern technologies is the scarcity of training data. This paper introduces a novel low-light video dataset, consisting of 40 scenes captured in various motion scenarios under two distinct low-lighting conditions, incorporating genuine noise and temporal artifacts. We provide fully registered ground truth data captured in normal light using a programmable motorized dolly, and subsequently, refine them via image-based post-processing to ensure the pixel-wise alignment of frames in different light levels. This paper also presents an exhaustive analysis of the low-light dataset, and demonstrates the extensive and representative nature of our dataset in the context of supervised learning. Our experimental results demonstrate the significance of fully registered video pairs in the development of low-light video enhancement methods and the need for comprehensive evaluation. Our dataset is available at DOI:10.21227/mzny-8c77.
Paper Structure (19 sections, 9 figures, 6 tables)

This paper contains 19 sections, 9 figures, 6 tables.

Table of Contents

  1. Introduction
  2. Existing low-light video datasets
  3. Real datasets
  4. Synthetic datasets
  5. State-of-the-art methods for video enhancement
  6. Low-light image enhancement
  7. Low-light video enhancement
  8. Video restoration
  9. Proposed low-light dataset
  10. Environment settings
  11. Ground truth generation
  12. The dataset
  13. PCDUNet framework
  14. Experiments
  15. Importance of fully registered data
  16. ...and 4 more sections

Figures (9)

  • Figure 1: Scene examples with varying light levels (10% and 20% low-light, 100% normal light) and different motion profiles (bottom row). From left to right columns: Mario3 with a static background, Book1, and Mario1. The length of the red lines across the normal-light frames represents the height of the x-t planes in the bottom row.
  • Figure 2: Examples of faulty SDSD video pairs, where the ground truth and the low-light frames are on the left and the right, respectively. (Top) The overexposed areas in the ground truth frame. (Bottom) The unaligned pair.
  • Figure 3: Setting environment showing Sony Alpha 7S II camera in 'angle' position, mounted on CineDrive system, with scene in the background.
  • Figure 4: Moving bunny scene with static background showing pixel value difference between the normal-light frame and the adjusted low-light frame before and after frame matching. Gray indicates zero error.
  • Figure 5: The proposed dataset: (From left to right) flowers, soft toy, wrapping, wood board, hats. (From top to bottom) Light levels of 10%, 20%, and normal light (100%).
  • ...and 4 more figures