Research library

Where Is the Bee? Detecting Tiny Pollinators with a Single Collaborative-Head Transformer

This paper presents the first-place single-model solution to the CVPPA@ECCV 2026 BuzzSpot Challenge, which localizes bees, bumblebees, hoverflies, and moths in high-resolution field video. The authors select Co-DINO with a Swin-L backbone after comparing ten detector configurations, then address severe class imbalance with rare-class crop mosaics and a class-weighted simplex equiangular tight frame loss. The resulting model reached 0.5062 mAP@[.5:.95] on the hidden FinalTest set without inference-time ensembling or test-time augmentation.

Publication details

Authors
Junsu Kim, Seungryul Baek
Organizations
🇰🇷 Ulsan National Institute of Science and Technology (UNIST)
Year
2026
Type
Preprint

Code and benchmark

Source and access

Relevancy to Gratheon

This work provides a strong, externally evaluated baseline for detecting bees when each insect occupies only a tiny fraction of a field image. Its comparison of detector families and its rare-class strategy are directly useful when Gratheon must distinguish honey bees from visually similar insects rather than merely detect generic motion.

The single-model result is operationally relevant because it avoids ensemble and test-time augmentation costs. The code, hidden-test score, and explicit treatment of class imbalance make the system a practical benchmark for crop-level pollination monitoring and for testing whether similar training methods improve small-bee detection around hive entrances.