Lightly
10 Case Studies
A Lightly Case Study
EUROCONTROL's aviation sustainability unit faced challenges in detecting and segmenting contrails from ash-RGB satellite imagery due to the domain shift from standard RGB, which made off-the-shelf models like YOLO and SAM underperform. Contributing factors included a limited dataset with a class imbalance and the poor utility of standard pretrained weights. This motivated the team to seek alternative representation learning and pretraining strategies with the help of Lightly.
Lightly implemented reproducible distillation workflows using its LightlyTrain product, which improved model performance. The solution led to an increase in true positives and upward shifts in metrics such as mAP and DICE for YOLO-based models. While comprehensive benchmarks are still being prepared, early findings indicate significant improvements, and the framework helped unify training configurations and simplify debugging efforts for the team.