SemiField Annotation Pipeline

 

Why

  • Labeling images for semantic segmentation is:

    • Expensive

    • Time consuming

    • Tedious

 

The burden of labeling weeds images has severely slowed the development of datasets necessary for utilizing artificial intelligence in agriculture, particularly in weed recognition, mapping and precision management applications.

 

The labeled pigweed segment pictured above in purple, was one of 62 in this cropped image. It took ~1.5 minutes to manually label. Multiply this by 60,000 images, and it becomes dreadfully obvious that manual labeling is not a scalable solution for developing a large image repository necessary for precision agriculture. At this pace, it would take 10.5 years to label a single years worth of images.

What

This project aims to reduce the burden of manually annotating images of weeds and other plants by implementing a semi-automatic annotation pipeline designed to iterate over batches of incoming images from 3 location around the US. Progressively more complex weed scenes are used to train, update, and apply detection and semantic segmentation models that assist labeling. The processing pipeline is iterative and implements a combination of photogrammetry, classical digital image processing, and deep learning techniques.

 

Photogrammetry

<image of orthomosaic>

Classical digital image processing

 

 

 

Deep Learning

<photo of bounding box detection or segmentation>

When

  • 3 years

  • summer for 6-12 weeks (weeds)

  • fall/winter 6-12 weeks (cover crops)

Who

Core developers include:

  • Matthew Kutugata

  • Chinmay Savadikar

  • Søren Skovsen

This project develops data for:

  • PSA network WIR

  • Researchers

  • Technology developers

  • and the precision agriculture community as a whole

Where

Processing occurs either on a PSA hosted Azure VM or locally at the Texas A&M high-performance computing weed tech lab.

The pipeline processes images collected in:

  • Raleigh, North Carolina (North Carolina State University)

  • Beltsville, Maryland (USDA-ARS research station)

  • College Station, Texas (Texas A&M University)

Github: https://github.com/precision-sustainable-ag/SemiF-AnnotationPipeline