aptemvs/mosquitoes-biodcase2026-task5

Mosquitoes — precomputed caches for BioDCASE 2026 Task 5 (CD-MSC)

Data backend for the Mosquitoes code repo (cross-domain mosquito species classification, BioDCASE 2026 Task 5). It holds the raw audio plus the derived caches the pipeline consumes, so a clone can reproduce every result without recomputing embeddings. The Hub layout mirrors the code repo; fetch_data.py pulls a group into place:

python fetch_data.py --group deployed   # ~3.7 GB — light probes + agreement gate
python fetch_data.py --group all        # HF caches (~58 GB) + Zenodo raw audio (~6.3 GB)

Contents

This HF dataset holds the derived caches (~58 GB). The raw audio is not re-hosted here — it is the official challenge dataset and is downloaded from Zenodo by fetch_data.py when you request --group heavy/all (see Provenance below).

group	size	files	what it is
deployed	3.7 GB	78	Perch / harmonic / bg-whitened embeddings (data/perch/*.npz), BirdMAE parquet, split metadata — runs the light probes + agreement gate
repro	9.5 GB	1,920	legacy/outputs + legacy/final (checkpoints, ensemble.json) — replays the historical leaderboard via import_runs.py / train_gate.py
heavy	19.6 GB	756	the log-mel streaming cache (data/feature) for the MTRCNN/EfficientAT path (raw audio comes from Zenodo)
extras	25.3 GB	14	additional foundation-model embeddings (Perch / sl-BEATs parquets, token tensors) used by the exploratory analysis/ scripts
raw_audio	6.3 GB	271,380	official Development_data.zip from Zenodo (not stored here); auto-fetched + extracted to data/raw_audio

Embeddings are frozen-encoder outputs (Google Perch, BirdMAE, sl-BEATs) plus hand-designed acoustic features (harmonic-comb, background-whitened spectrum); they carry no model weights.

Provenance & attribution

The audio and labels are redistributed from the official challenge dataset under its CC BY 4.0 license; all embeddings/features here are derived from it:

BioDCASE 2026 Challenge: Cross-Domain Mosquito Species Classification. Yuanbo Hou, Vanja Zdravkovic, Marianne Sinka, Yunpeng Li, Kathy Willis, Stephen Roberts. Zenodo, 2026. DOI 10.5281/zenodo.20478577. Licensed CC BY 4.0.

Challenge task page: https://biodcase.github.io/challenge2026/task5. Challenge coordinators: Yuanbo Hou, Vanja Zdravkovic, Marianne Sinka, Kathy Willis, Stephen Roberts (University of Oxford); Yunpeng Li, Mark Plumbley (King's College London); Wenwu Wang (University of Surrey).

The underlying dataset: 9 mosquito species across 5 acquisition domains (location / device / acoustic environment), 271,380 clips (~60.7 h), highly imbalanced. The metric is BA_unseen — species-balanced accuracy on the domain each species is not trained on.

Citation

If you use these caches, please cite the source dataset and the challenge papers:

@dataset{hou2026cdmsc_dataset,
  title     = {{BioDCASE} 2026 Challenge: Cross-Domain Mosquito Species Classification},
  author    = {Hou, Yuanbo and Zdravkovic, Vanja and Sinka, Marianne and
               Li, Yunpeng and Willis, Kathy and Roberts, Stephen},
  year      = {2026},
  publisher = {Zenodo},
  doi       = {10.5281/zenodo.20478577},
  url       = {https://doi.org/10.5281/zenodo.20478577}
}

@article{hou2026biodcase_baseline,
  title   = {{BioDCASE} 2026 Challenge Baseline for Cross-Domain Mosquito
             Species Classification},
  author  = {Hou, Yuanbo and others},
  journal = {arXiv preprint arXiv:2603.20118},
  year    = {2026}
}

@inproceedings{hou2026drbiol,
  title     = {Learning Domain-Robust Bioacoustic Representations for Mosquito
               Species Classification with Contrastive Learning and Distribution Alignment},
  author    = {Hou, Yuanbo and others},
  booktitle = {ICASSP 2026 - IEEE International Conference on Acoustics, Speech
               and Signal Processing (ICASSP)},
  pages     = {15207--15211},
  year      = {2026},
  doi       = {10.1109/ICASSP55912.2026.11464393}
}

@inproceedings{hou2025mtrcnn,
  title     = {Sound-Based Recognition of Touch Gestures and Emotions for
               Enhanced Human-Robot Interaction},
  author    = {Hou, Yuanbo and Ren, Q. and Wang, Wenwu and Botteldooren, D.},
  booktitle = {ICASSP 2025 - IEEE International Conference on Acoustics, Speech
               and Signal Processing (ICASSP)},
  pages     = {1--5},
  year      = {2025},
  doi       = {10.1109/ICASSP49660.2025.10890031}
}

License

CC BY 4.0, inherited from the source dataset. You may share and adapt with attribution to the dataset authors above. Note the challenge rules forbid external labelled mosquito data for official submissions — the derived features here are computed only from the official dataset.