I have found two meanings of how "ground truth" is being used in machine learning:

  • Something to be assumed true
  • Something previously validated as true

Although similar in detail the two can differ. Maybe they are used interchangeably, but a citation would be nice for any meaning. I have not been able to find a definition in literature only seeing it being used in papers.

  • 1
    $\begingroup$ I'm only familiar with informal, colloquial uses, but given the diversity of specific definitions mentioned here, you may have more luck if you specify a discipline. $\endgroup$ – Sean Easter Aug 10 '15 at 14:13

The term "ground truth" was coined in the geological/earth sciences to describe validation of data by going out in the field and checking "on the ground". It has been adopted in other fields to express the notion of data that is "known" to be correct. In my personal experience it is widely used in biometrics and computer vision. The term "ground truth error" is also in wide use, illustrating the fact that what we "know" is not always correct.


@article {Dictionary.com2015, title = {Dictionary.com's 21st Century Lexicon}, month = {Aug}, day = {18}, year = {2015}, url = {http://dictionary.reference.com/browse/ground truth}, }

for an online definition.

See @book{krig2014computer, title={Computer Vision Metrics: Survey, Taxonomy, and Analysis}, author={Krig, Scott}, year={2014}, publisher={Apress} } Chapter 7, "Ground Truth Data, Content, Metrics and Analysis" for a discussion of ground truth in Computer Vision -- available in print and eBook formats.

There is an interesting blog at thegroundtruthproject.org

NASA has a glossary of term that includes ground truth -- see http://podaac.jpl.nasa.gov/Glossary

  • $\begingroup$ Do you have the link to the thegroundtruthproject.org blog post? Thx $\endgroup$ – herve Apr 24 '17 at 9:38

In most cases it is used as 'real true' e.g. scikit in Python, with specific examples such as image recognition eg Nunez-Iglesias et al. PLOS One, character recognition Luis von Ahn et al. Science.

But how close is the 'real true' to a fixed value can depend on the complexity of input and whether "reference data can be less accurate than recognition system being evaluated" Lopresti and Nagy and a search for ground-truthing issues could yield further results eg this overview of symbol recognition. (Whereas assumed vs. validated would refer largely to a particular hypothesis/implementation.)


This is not exactly a definition, but a brief, nuanced description of ground truth in machine learning by James Kobielus at IBM: http://www.ibmbigdatahub.com/blog/ground-truth-agile-machine-learning

Within machine learning, I would call ground truth a human-defined truth or an external truth rather than an epistemological truth or actual truth. Ground truth is the foundation of supervised machine learning.


Meaning “according to the “assumed””. The assumed is the ground, as taken the only referral voltage in electrical systems. The term was adopted by computer scientists working in machine learning who were inherited from electrical engineers.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.