Ground Truth

The real data treated as the “correct answer” for training and evaluating a model.

Ground truth originally comes from remote sensing, where it referred to data collected on the ground to calibrate satellite measurements; machine learning and statistical modeling borrowed the term to mean whatever data is treated as the correct answer during training and evaluation, such as an image's class label or an object's true position. Ground truth isn't necessarily perfect — human annotators make mistakes and sensors have error — so data quality directly caps how good a model can get. In embodied AI, imitation learning treats the actions recorded during human teleoperation as the action ground truth; a simulator can read out an object's exact pose, contact forces, and other precise state directly, so it's often used as ground truth for training a perception module, or as privileged information for a teacher policy; real-robot experiments often use poses measured by a motion-capture system as ground truth for evaluating an algorithm.

ExampleWhen training a 6D pose-estimation model on simulator-rendered images, each image's true object pose comes straight from the simulator — that's ground truth, and it needs no human labeling.

Also called

GT, Gold Label

Related

Data Annotation · Privileged Information · Supervised Learning · Action Label · Synthetic Data · Motion Capture

Sources

Wikipedia: Ground truth
Google Machine Learning Glossary