Autolabelign with Lightly works in four simple steps

1. We look at the unlabeled data and select a small subset using Lightly's Active Learning algorithms
2. The labeling partner labels the subset of images
3. We use the labeled data to train a model on the labeled data using few-shot learning
4. We create predictions for the unlabeled data

In the next iteration, we decide which subset of images should go to the next labeling batch and which subst of images is likely to be predicted correctly (autolabeled). The autolabeled data can optionally go through a QA process of the labeling partner to guarantee correctness.

Different Autolabeling Techniques Explained

There are different ways f how to autolabel a dataset. The two main approaches are either using few-shot learning or using prior information and physical rules into account.

Few-Shot Learning

Few-Shot Learning is a research field that aims to train machine learning models with very few labeled images. Often 3-5 images per class can be enough to kick off the process.

Rule-Based Labeling

When working with videos we can most often assume linear motion between consecutive frames. We basically assume objects do not suddenly jump of out of the frame and are visible throughout several frames. Machine Learning models, however often create flickering when working with videos, as a slight variation in the input data can drastically affect the prediction confidence. We can make use of that using model assertions to flag images where the prediction was flickering.