Featured study · SR-PLNT-005
Proposal vs. Transformer
How do Faster R-CNN, Swin Transformer, and DETR compare across small, medium, and large object scales for weed detection in morphologically similar green crops?
The study evaluates where each architecture succeeds and fails in dense, visually ambiguous crop scenes. Faster R-CNN led overall and was the only model to retain meaningful precision on small objects at IoU 0.5.
- field images collected
- 2,600
- images in evaluation set
- 1,052
- custom box annotations
- 7,000
- best overall mAP
- 76.1%
research workflow
- Data collection
- Manual annotation
- 80 / 10 / 10 split
- Model training
- SAM inference
- COCO evaluation
test results · IoU 0.5
Architecture comparison
AP small = objects ≤ 32² px
Faster R-CNN
Best overall and strongest on dense spinach canopies
- mAP
- 76.1%
- mAR
- 78.8%
- AP small
- 15.8%
Swin Transformer
Best on structured lettuce fields at 99.6% AP
- mAP
- 65.3%
- mAR
- 70%
- AP small
- 1%
DETR
Struggled with crowded scenes and overlapping plants
- mAP
- 47.8%
- mAR
- 58.8%
- AP small
- 0.7%