Auto-Labeling Image Data for Aerospace Autonomy

Author: Nic Ceccanti

Image: VTI's cameras capture live runway footage during approach.

Many aerospace companies still treat labeling as a slow external dependency: export data, wait for labels, inspect the results later, then clean up problems after the fact. VTI’s approach is different. We build labeling into the autonomy development process so that data collection, annotation, review, and model improvement move together.

Modern aerospace autonomy depends on good ground truth data. Cameras and onboard sensors can record a lot of real-world imagery, but that data is not useful for ML until the important features in the image are labeled. For vision models, this usually means labels that allow the model to detect and localize the things that matter during flight.

At VTI, we use a semi-automated image labeling workflow. The goal is to automate as much of the labeling process as possible while keeping the data quality bar high. Today, a small team of labelers and reviewers works across the dataset, using the tooling to produce and verify labels before the data is used for training or evaluation.

This is important because aerospace images are not generic computer vision data. Lighting, runway geometry, aircraft attitude, sensor placement, distance from the runway, and weather all change what the model sees. A label that looks “close enough” in a generic dataset may not be good enough for a model that is learning features that are used in localizing the aircraft.

A big advantage of our approach is that labeling is connected to the engineering workflow. The labels are not treated as a one-off deliverable from an outside vendor. They are tied back to the data, the flight phase, the review process, and the model behavior we are trying to improve. This makes it easier to understand what data we have, what data still needs review, and where the model may need more examples.

We also keep the labeling work close to the team building the system. Rather than sending image data to a large external labeling operation, we use a focused U.S.-based team with aviation experience. That matters because the reviewers understand the operational context better than a generic labeling workforce. They are closer to the engineers, closer to the flight test data, and better positioned to catch issues and notify engineers.

The result is a better feedback loop. Automation helps us scale across more imagery. Human reviewers focus on ambiguity, consistency, and label quality. Engineers get datasets that are more reliable and easier to reason about when training and evaluating models.

The long-term direction is to keep pushing more of this workflow into automation. The advantage today comes from using automation aggressively while keeping experienced reviewers focused only on the parts where human judgment still improves dataset quality. For aviation autonomy, that is the difference between having a pile of images and having training data that can actually improve model performance.