Interactive Segmentation With Prototype Learning for Few-Shot Root Annotation

Xiaolei Guo, Alina Zare, Lisa Anthony, Felix B. Fritschi

Abstract: Fine-scale pixel-level annotation of minirhizotron root images is a less common and challenging task. We present an interactive segmentation framework to accelerate root annotation. We leverage the concept of few-shot segmentation so that the pretrained model can be effectively fine-tuned and transferred to an unseen category. To provide immediate feedback for real-time interaction, we adapted a UNet architecture by attaching lightweight embedding layers which leveraged a prototype learning (PL) approach to efficiently learn the data metric in the embedding space. The prototypes optimized by the prototype loss preserve the within-class data variation, enabling effective fine-tuning. Furthermore, we designed a system with our interactive annotation framework and experimented with real users to validate the approach.

External IDs:doi:10.1109/tgrs.2025.3556799