Efficient Training of Object Detection Models for Automated Target Recognition

摘要

GATR™ (Globally-scalable Automated Target Recognition) is a software system for object detection and classification on a worldwide basis developed by Lockheed Martin. One of the targets it detects are oil/gas fracking wells. In this work we explore the following techniques for efficiently training a fracking well detector in WorldView satellite imagery: determining the optimal ground sample distance (GSD) of data to use for training/inference, leveraging the use of auto-generated training labels, training on different sized subsets of the training data, and using active learning to determine the best subsets of data to train with. We compare the mean average precision (mAP) for models trained at different GSDs from 0.3 to 12 m and find optimal performance at 1.2 m GSD (mAP = 0.91), but acceptable performance all the way to 6 m GSD (mAP = 0.80). We also show that using auto-generated training labels from fracking well permits results in a model with mAP reduced by less than 4% (IoU threshold = 0.2) compared to a model trained on the human annotated dataset. We show that training on just 20% of the full dataset results in a mAP drop of less than 6% (mAP = 0.86). Finally, we compare multiple active learning selection methods.