BeeHunters

Official implementation of Where Is the Bee? Detecting Tiny Pollinators with a Single Collaborative-Head Transformer.

This repository reproduces our 1st-place single-model solution to the BuzzSpot Challenge at CVPPA@ECCV 2026. The final checkpoint achieved 0.5061609965 mAP@[.5:.95] on FinalTest.

Method

BeeHunters uses Co-DINO with a Swin-L backbone, initialized from the Objects365-to-COCO checkpoint, and adds an auxiliary class-weighted fixed-simplex ETF loss adapted from NC-FSCIL on matched positive decoder queries. Training follows three phases:

Phase	Training source	Epochs	Learning rate
1	Combined train+valid keyframes	12	1e-4
2	Class-aware crop mosaics	3	1e-5
3	Low-rate fine-tuning on original keyframes	2	5e-6

Setup

The recorded environment uses Python 3.10 and CUDA 11.8. Run all commands from the repository root.

bash scripts/setup_env.sh
source .venv/bin/activate
bash scripts/download_checkpoints.sh pretrained

Download and prepare the BuzzSpot data:

bash scripts/download_data.sh
bash scripts/prepare_data.sh
python scripts/build_cropbank.py

Training

Run the clean three-phase recipe in order on four NVIDIA A100 80 GB GPUs.

1. Phase 1: Full-data training (12 epochs). Start from the pretrained Co-DINO checkpoint and train on the combined train and validation keyframes.

   bash scripts/train.sh 1 0,1,2,3

2. Phase 2: Class-aware crop-mosaic fine-tuning (3 epochs). Resume Phase 1 and fine-tune on mosaics assembled from object-centered crops.

   bash scripts/train.sh 2 0,1,2,3

3. Phase 3: Low-rate fine-tuning on original keyframes (2 epochs). Resume Phase 2 and fine-tune on the original train and validation keyframes with a 5e-6 learning rate.

   bash scripts/train.sh 3 0,1,2,3

The scored checkpoint's final training epoch resumed on three GPUs after an interruption. This runtime provenance and the clean replay recipe are separated in docs/REPRODUCTION.md.

Inference

Download the final checkpoint and create the FinalTest submission:

FINAL_CHECKPOINT_URL='https://github.com/jjunsss/BeeHunters/releases/download/v1.0.0/buzzspot_codino_etf_final_inference.pth' \
  bash scripts/download_checkpoints.sh final
python scripts/infer.py --gpus 0,1,2

Acknowledgements

Thanks to the following open-source projects for making this work possible:

• MMDetection
• CO-DETR
• BuzzSpot DevKit