Pre-labeling is the practice of running a model over a dataset to generate initial labels in batch before humans review them, so annotators start from a pre-filled draft instead of a blank image. It front-loads the model's work and turns labeling into correction.
Pre-labeling, also called pre-annotation, runs a model across your unlabeled data ahead of time and writes its predictions in as a first-pass set of labels. When annotators open a sample, the boxes or masks are already there, and their job is to fix what is wrong, add what is missing, and remove false positives. Unlike model-assisted labeling, which suggests interactively as you work, pre-labeling is a batch step done before human review even begins. It is one of the most common ways teams cut labeling time.
Key takeaways
A model labels the whole dataset in batch first, humans then correct the draft.
It is distinct from model-assisted labeling, which is interactive, pre-labeling happens upfront.
It works best when the pre-labeling model is reasonably close to your data, otherwise correction costs more than starting fresh.
Pre-labeling vs related approaches
How pre-labeling compares to the other model-driven labeling approaches.
How pre-labeling compares to the other model-driven labeling approaches.
Approach
When the model runs
Human starts from
Manual (from scratch)
Not at all
A blank sample
Pre-labeling
Batch, before review
A pre-filled draft
Model-assisted
Interactively, during labeling
A blank sample plus live suggestions
Agentic
A background job, guided by a trained agent
The agent's labels, then review
Fully automatic
Batch, no review
Nothing, labels used as-is
How it works, and how FiftyOne fits
You run a model over the dataset, store its predictions as labels, and send those to annotators as the starting point. In FiftyOne this is a natural fit: apply a model to a dataset, save the predictions, review which pre-labels are good enough to keep, and route the rest for correction before training.
Why it matters
Pre-labeling can cut labeling effort dramatically when the model is decent, because correcting beats creating. There is a break-even point people miss. If the pre-labels are bad enough, fixing them, deleting wrong boxes, repositioning, undoing the model's mistakes, costs more than labeling from scratch, and annotators anchor on the model's errors and miss what it missed entirely. Pre-labeling pays off above a quality threshold and backfires below it, which is why you validate the pre-labeling model on a sample before running it across everything.
Frequently asked questions
What is the difference between pre-labeling and model-assisted labeling?
Pre-labeling fills in labels in batch beforehand, model-assisted suggests interactively while you work.
Is pre-labeling the same as auto-labeling?
Pre-labeling is a form of auto-labeling aimed at giving humans a starting draft to correct.
When does pre-labeling backfire?
When the model is poor, correcting its mistakes costs more than labeling fresh and biases annotators toward its errors.