A dataset for when your camera sees better than you do

What it does ExDark is a computer vision dataset of low-light images spanning 10 conditions from very low-light to twilight. It includes 12 object classes (PASCAL VOC-style) with both image-level labels and bounding box annotations. The authors also provide MATLAB code for low-light image enhancement in the SPIC folder.

The interesting bit Most vision datasets assume decent lighting. ExDark doesn’t — it targets the awkward gap where standard detectors start guessing. The 10 condition gradation is the useful detail: you can benchmark how badly your model fails as light drops, rather than treating “dark” as a single bucket.

Key highlights

• 7,363 images across 10 low-light conditions
• 12 object classes with image-level and bounding-box annotations
• Published in CVIU 2019; BSD-3 licensed (commercial use requires contact)
• Includes low-light enhancement code (SPIC folder, MATLAB)
• Companion paper: “Getting to Know Low-light Images with The Exclusively Dark Dataset”

Caveats

• MATLAB-only for the enhancement code; no Python reimplementation mentioned
• Dataset link was updated in 2022 — verify current accessibility
• Commercial licensing requires explicit contact rather than straightforward open use

Verdict Worth bookmarking if you’re training detectors for night cameras, drones, or automotive vision. Skip if you need a plug-and-play Python pipeline today — you’ll be translating MATLAB or waiting for community ports.