IMAGE
COMMERCIAL
LABELED
Find a surface with at least ten countable objects, photograph it from above, count carefully, enter the exact number. The count is checked on review, and images with fewer than ten objects, heavy occlusion or a wrong count are rejected; ten photos per contributor at most, worldwide, so the class mix is wide. The count is the only label; boxes or dots can be added cheaply because the count tells annotators when they are done. Counting is a perceptual task VLMs still get wrong: a December 2025 evaluation on FSC-147 (arXiv 2512.15254) put the best model near 0.5 exact-count accuracy, and FSC-147 itself is 6,000 images under unstated terms.
Volume
25,000+
Images
Contributors
5,000+
Regions
100+ countries
Updated
What’s in the collection and how it’s delivered.
Data modality
Images
Annotated
Labeled (human-verified object count)
Exclusivity
Non-exclusive rights
Capture tool
Phone camera (in app)
Use cases
Object counting and class-agnostic counting evaluation, dense detection in clutter, VLM numeracy benchmarks
Format
JPG / PNG / HEIC
Resolution
Native phone camera resolution (JPG, PNG, HEIC)
Total volume
25,000+ images (approx. 75 GB)
Regions
100+ countries
Purity level
99% passed QA review
Where this data comes from and how it’s licensed.
Every file comes straight from the contributor who provided it and is rights-cleared before it’s listed, so your license traces back to the source.
Contributor
Provides the data and uploads it straight to Kled.
Kled
Clears the rights before anything is listed.
Your team
Licenses it under the terms below.
License terms
Non-exclusive commercial training license.
Consent
Each contributor opted in to this use when they uploaded.
Specs reflect the current collection and may change as new data is added. Final terms are set in your license agreement.
