What is semi-ground truth and how it is different from ground truth?

In machine learning and computer vision, ground truth refers to the correct or expected output for a given input, which is often obtained through manual annotation or labeling. On the other hand, semi-ground truth is a partially annotated version of the ground truth, where some portions of the data have been labeled, but the remainder has not.

Semi-ground truth is commonly used in scenarios where the full ground truth is too expensive or time-consuming to obtain. Instead, a subset of the data is labeled, which can be used to train or evaluate models. However, since only a portion of the data is labeled, the performance of the models is limited by the availability and quality of the labeled data.

For example, let's say you want to train a machine-learning model to detect objects in images. You have a dataset of 10,000 images, but labeling all of them with ground truth object annotations is too time-consuming and expensive. Instead, you decide to label a random subset of 1,000 images with object annotations and use this semi-ground truth data to train the model. The model will learn to detect objects based on the available annotations, but its performance on the remaining 9,000 unlabeled images will be limited by the quality and coverage of the semi-ground truth data.

Semi-ground truth can also be used in combination with other data augmentation techniques, such as synthetic data generation or transfer learning, to further improve the performance of the models.