Deep Learning-Based Detector and Automatic Labeler for Vehicles in Overhead Imagery

About this technology

Overview

We want to detect vehicles in a series of overhead images to ultimately track these vehicles. Off-the-shelf detection models are not good at detecting vehicles in overhead imagery. Open source labeled dataset COWC (Cars Overhead with Context) provides center point annotations, but we want bounding boxes to train a better model.We develop a deep learning pipeline to automatically convert center point annotations into bounding box labels using SAM segmentation model and Grounding DINO feature embeddings. Using these labels, we train a deep learning model to detect vehicles at varying resolutions.