What is pseudo-labeling in detection?

Pseudo-labeling is a semi-automatic learning method that uses unlabeled data in object detection tasks. The idea is that a model trained on labeled data predicts labels for unlabeled images. These predicted labels are called pseudo-labels.

Then, the model is further trained on a combined dataset: labeled data + unlabeled data with pseudo-labels. This helps to utilize more data and improve detection quality, especially when labeled data is scarce.

It is important to control the quality of pseudo-labels, for example, by discarding predictions with low confidence to avoid degrading the training.