MAVNet is a compact neural network designed to run semantic segmentation on micro aerial vehicles with limited onboard computing resources. Its results illustrate how model design, task-specific datasets and evaluation metrics shape the trade-off between speed and accuracy.