FollowMe: Automating Precise Image Annotation Across Various Environments

Abstract

Precise image annotation is essential for training advanced computer vision models. However, making high-accuracy, pixel-level image annotation efficient is never easy: manual labeling is highly accurate but time-consuming, while automatic labeling that using state-of-the-art models is much faster but often yields unsatisfactory errors. In this work, we propose FollowMe, which combines the strengths of both approaches by minimizing human effort without sacrificing annotation accuracy for images captured across different environments (e.g., different times, seasons, and weather conditions) of the same physical scene. FollowMe leverages high-quality annotations from one image to automatically annotate "similar" images under other conditions, requiring only lightweight human review and correction. We evaluate FollowMe using one public dataset released in our prior work [1]. FollowMe effectively reduces human efforts by 94% while maintaining an average F1 score of 0.9.