Sample frames for this project are available under NDA. Write to bilal@wortel.ai and we will share the relevant before and after pairs.
Challenges
- Drone-captured aerial imagery showed trees only from a top-down perspective.
- Tree shadows closely resembled actual trees and had to be excluded.
- Separating target almond trees from nearby trees and vegetation proved difficult.
- Rows of trees placed multiple objects very close together.
- Soil, vegetation, shadows and other objects made the aerial imagery visually complex.
- Variations in tree size, position and appearance complicated accurate box placement.
Related work
Ground Cover Quadrat Segmentation
2,275 cover-crop plot images annotated in Labelbox: plot corners boxed, then every pixel classed as groundcover, weed, residue, crop or soil.
Apple Detection
Bounding boxes around individual apples on tree branches, where dense leaves and occlusion hid much of each fruit.
Broccoli Stand Count
Bounding boxes on individual broccoli plants for automated stand counting, separating closely packed plants from surrounding weeds.