Crocos-V1: Enhancing Mask Leakage and Bounding Box Localization for Real-Time Crop/Weed Instance Segmentation*

Abstract

Translate Article

This paper exposes a new algorithm designed to enhance the real-time crop/weed instance segmentation. The approach combines a learning-based method for instance segmentation with a feature model-based image processing strategy that leverages the vegetation characteristics of crops. The proposed algorithm compensates for the shortcomings of each method performing independently. The image processing strategy achieves precise crop segmentation by generating finely refined masks; but may introduce errors in weed segmentation. Conversely, instance segmentation methods perform well in accurately identifying both crops and weeds, but can produce imperfect masks in the presence of inaccurate bounding boxes. The experiments are conducted in different evaluation campaigns including the ACRE international competition framework. The results demonstrate that integrating color feature segmentation with state-of-the-art instance segmentation methods improves overall segmentation accuracy, achieving up to 0.80 mAP for maize crops and 0.83 mAP for bean crops, while maintaining real-time computational efficiency.