Skip to content
Dataset By Humans
Computer vision & video

DPED dataset and alternatives for smartphone photo enhancement

What the DPED dataset contains, other photo enhancement and low-light datasets researchers use, and what to look for in a commercially licensed alternative.

By Dataset By HumansPublished 2 min read

In this article

Short answer: DPED (the DSLR Photo Enhancement Dataset) pairs photos from older smartphones with photos of the same scenes taken on a DSLR, so a model can learn to make phone photos look DSLR-quality. It remains a standard research benchmark, but its phones are more than a decade old and academic sets like it are generally released for research, not commercial products. For production work you will usually want current-phone captures with a commercial license.

What DPED is

DPED was introduced by Ignatov et al. in "DSLR-Quality Photos on Mobile Devices with Deep Convolutional Networks" (ICCV 2017). The authors captured the same scenes simultaneously with three phones (iPhone 3GS, BlackBerry Passport and Sony Xperia Z) and a Canon 70D DSLR, then aligned the images into training patches. It made photo enhancement a learnable, measurable task and is still widely cited.

Other datasets researchers use

DatasetWhat it containsTypical use
DPEDPhone and DSLR photos of the same scenesPhone-to-DSLR enhancement
MIT-Adobe FiveK5,000 raw photos, each retouched by five expertsLearned photo retouching
SID ("Learning to See in the Dark")Short- and long-exposure raw pairs in low lightExtreme low-light imaging
LOLPaired low-light and normal-light imagesLow-light enhancement
Zurich RAW to RGBPhone raw images paired with DSLR imagesLearned ISP pipelines

Always check each dataset's license before using it in a product; terms differ and many are non-commercial.

Why older benchmarks fall short for products

  • Old sensors. Modern phones use multi-frame HDR, night modes and heavy computational processing that a 2013-era phone never had.
  • Few conditions. Production failures come from rain, backlight, neon signs and mixed lighting, which small benchmarks cover thinly.
  • Narrow geography. Scenes, signage and lighting from one or two cities do not represent where your users live.
  • Licensing. Research-only terms do not cover a shipped camera feature or a commercial API.

What to look for in a commercial alternative

  1. Captures from the phones your users carry today, with EXIF kept (device, ISO, exposure, focal length)
  2. Paired or bracketed shots of the same scene across camera settings, if you train enhancement models
  3. Real adverse conditions: night, rain, fog, glare, backlight
  4. Privacy processing: faces and plates blurred, GPS removed
  5. A written commercial license and a datasheet

Our low-light and adverse-weather collection

Our collectors photograph night streets, rain, fog and backlit scenes on current smartphones, including paired shots across camera settings, and every image is reviewed by a person. See our curated photo datasets or request a collection to your spec.

dped datasetimage enhancementlow light datasetsmartphone photos