Ensemble-Based Image Annotation for Real-Time Human Detection
Alexander Filonenko, I. Yu. Popov, Kang-Hyun Jo · 2024
This paper introduces an ensemble-based image annotation framework for human detection, designed to optimize real-time performance of the training of a lightweight neural network. By synergistically combining the strengths of multiple models, our approach achieves substantial enhancements in cross-dataset generalizability and annotation accuracy, surpassing manual annotation in specific scenarios. To facilitate widespread adoption, we provide an open-source library and publicly available pretrained models for ensemble-based image annotation in real-time human detection applications. Notably, one of the constituent models in the ensemble has secured a top-2 ranking in the CrowdHuman dataset leaderboard.