Ground truth
lastmod 2022-08-05created 2022-08-042 edits
Ground Truth is information known to be real or true, provided by direct observation and measurement rather than by inference.
Interpret it using whichever of the explanations below makes more sense to you.
- Ground truth is the labeled data used to train and test AI models.
- Ground truth is a term for real-world data used to train and test the output of AI models.
- A term for the actual true value that serves as the reference for comparison when you want to evaluate how accurate some (computed, estimated, predicted) value is.
- In machine learning, the term "ground truth" refers to the accuracy of the training set's classification for supervised learning techniques.
Etymology
The Oxford English Dictionary (s.v. "ground truth") records the use of the word "Groundtruth" in the sense of "fundamental truth" in Henry Ellison's poem "The Siberian Exile's Tale", published in 1833.
Statistics and machine learning
"Ground truth" can be seen as a conceptual term relating to knowledge of the truth concerning a specific question. It is the ideal expected result. It is used in statistical models to prove or disprove research hypotheses.
The term "ground truthing" refers to the process of gathering the proper objective (provable) data for this test.
Label vs Ground-truth
- The meaning of "Ground Truth" as used in machine learning/AI
- A "label", as its other name "answer key" suggests, is a value whose answer is clearly fixed.
- "ground-truth" is 'the answer we decided on', 'the answer we hope our model will predict as the answer we want'.
About
Ground-truth is a term that originated in meteorology and means information collected at a particular location. Ground-truth is usually interpreted as 'ground survey information'. When observing the Earth from far away, as with satellites, you get a wide view of the Earth as a whole, but seeing the actual ground structure in detail is difficult because light passing through clouds and the atmosphere distorts the real picture.
From a machine learning perspective, ground-truth is used to denote the original or actual values of the data to be learned. Let's look at an example of an algorithm that aims at object detection in images.
Background-removal-with-deep-learning.png
In the figure above, the image on the far right (->) is the ground-truth, showing the information of the objects in the actual image.
We can infer that green is the bicycle and pink is the person riding the bicycle.
The pictures on the left are the results of object detection on the photo; they do not fully match the ground-truth image, but you can see that a deep learning model trained by machine learning confirms the presence of objects at those locations, and they match the ground-truth to some degree.
Next, let's look at a photo in which the locations of detected objects are represented with bounding boxes when performing object detection on an image.
Iou_stop_sign.png
The photo above shows a detected sign represented by bounding boxes: the blue box is the ground-truth and the red box is the area predicted and marked by the deep learning model. For a deep learning model that detects objects well, a prediction whose bounding box is closest to the ground-truth indicates a model with good performance.
So by what criterion can we tell whether a deep learning model predicted well? To compare this numerically, we use the IoU (Intersection over Union) formula.