Active learning with point supervision for cost-effective panicle detection in cereal crops

Area of Science:

• Agricultural Science
• Computer Vision
• Machine Learning

Background:

• Panicle density is crucial for estimating cereal crop yield (wheat, sorghum).
• Accurate phenotyping requires computer vision-based object detection.
• Deep learning models need extensive bounding-box labeled data, which is costly and time-consuming to acquire for diverse crops.

Purpose of the Study:

• To develop a cost-effective method for training reliable panicle detectors for cereal crops.
• To reduce the expense and time associated with acquiring large labeled image datasets for crop phenotyping.
• To enable widespread adoption of automated object detection in agriculture.

Main Methods:

• Proposed a point supervision-based active learning approach for panicle detection.
• Implemented an iterative querying strategy for informative images, interacting with a human annotator.
• Utilized low-cost weak labels (object centers) to identify images requiring expensive strong labels (bounding boxes).

Main Results:

• Demonstrated promising results on Sorghum and Wheat datasets.
• Achieved over 55% savings in labeling time on Sorghum compared to baseline methods.
• Achieved over 50% savings in labeling time on Wheat compared to baseline methods.

Conclusions:

• Developed a cost-effective method for training reliable cereal crop panicle detectors.
• The method offers significant benefits for plant breeders and agronomists in crop yield estimation and management.
• Facilitates real-time visual crop analysis for research on crop responses to experimental conditions.