Low-light Object Detection (LOD) Dataset

Dataset Overview

Dataset Name

• Low‑light Object Detection (LOD) Dataset

Dataset Source

• Proposed by Yang Hong, Kaixuan Wei, Linwei Chen, and Ying Fu at BMVC 2021.

Dataset Content

1. Image Types

   • RGB‑normal: long‑exposure normal‑lighting sRGB images.
   • RGB‑dark: short‑exposure low‑lighting sRGB images.
   • RAW‑normal: long‑exposure normal‑lighting RAW images.
   • RAW‑dark: short‑exposure low‑lighting RAW images.

2. Annotation Files

   • Each image is accompanied by an .xml annotation file (for all four image types).
   • Updated Dec 2021: an additional .xml file based on RAW‑normal annotations is provided.

Dataset Download

• All files are available via Baidu Cloud; extraction code: “2021”.

Dataset Characteristics

• Every short‑exposure image has a corresponding long‑exposure image as ground truth.
• Captured with a Canon EOS 5D Mark IV; raw sensor data is provided.
• Filenames are pure numbers; short‑exposure filenames equal long‑exposure filenames + 1.

Citation

• If you use this dataset, please cite:
  • Yang Hong, Kaixuan Wei, Linwei Chen, and Ying Fu, “Crafting Object Detection in Very Low Light”, BMVC 2021.