YOLO-Guided Pretraining for Apple Counting under Limited Annotations

Giuliano Ramirez, Orietta Nicolis, Hans Lobel, Billy Peralta

Keywords: Apple Counting, YOLO, Pretraining

Abstract:

Precision agriculture requires scalable and data-efficient fruit
counting, as manual tallies are slow and vision-based models still rely heavily on fully labeled datasets. We introduce
a pseudo/semi-supervised pipeline where a YOLO detector generates pseudo-labels to pretrain a U-Net segmenter.
Specifically, we treat YOLO as a teacher to pretrain a specialized counting model, transferring detection cues into a
segmentation-based counter. On the MinneApple benchmark,
this approach reduces average error from RMSE 15.35 and
MAE 11.50 to 13.03 and 10.10, while raising R2
from 0.66 to 0.77, with stable behavior around detector thresholds of 0.065–0.075. When labeled data are reduced (e.g.,
from 67/13/10 to 17/3/10 splits), the pretrained model consistently outperforms the baseline, increasing R2
by more than 0.12 in the most label-scarce regime. Scaling pseudo-labeled
pretraining while reducing supervised data by 25–75% further improves performance, with R2
gains up to 0.71 compared to 0.56. These findings demonstrate that detector-driven
pretraining is an effective and label-efficient strategy for
fruit counting under data scarcity, and point toward extensions across crop types, sensing modalities, and deploymentoriented advances such as adaptive thresholding, active learning, and domain adaptation.

Submission Number: 5