AI-EcoNet/HUGO-Bench-Paper-Reproducibility

HUGO-Bench Paper Reproducibility

Supplementary data and reproducibility materials for the paper:

Vision Transformers for Zero-Shot Clustering of Animal Images: A Comparative Benchmarking Study

Hugo Markoff, Stefan Hein Bengtson, Michael Ørsted

Aalborg University, Denmark

Dataset Description

This repository contains complete experimental results, pre-computed embeddings, and execution logs from our comprehensive benchmarking study evaluating Vision Transformer models for zero-shot clustering of wildlife camera trap images.

Related Resources

• Source Images: AI-EcoNet/HUGO-Bench - 139,111 wildlife images
• Code Repository: Coming soon

Repository Structure

├── primary_benchmarking/          # Main benchmark results (27,600 configurations)
├── model_comparison/              # Cross-model comparisons
├── dimensionality_reduction/      # UMAP/t-SNE/PCA analysis
├── clustering_supervised/         # Supervised clustering metrics
├── clustering_unsupervised/       # Unsupervised clustering results
├── cluster_count_prediction/      # Optimal cluster count analysis
├── intra_species_variation/       # Within-species cluster analysis
│   ├── train-*.parquet           # Analysis results
│   └── cluster_image_mappings.json  # Image-to-cluster assignments
├── scaling_tests/                 # Sample size scaling experiments
├── uneven_distribution/           # Class imbalance experiments
├── subsample_definitions/         # Reproducible subsample definitions
├── embeddings_*/                  # Pre-computed embeddings (5 models)
│   ├── embeddings_dinov3_vith16plus/  # 120K embeddings, 1280-dim
│   ├── embeddings_dinov2_vitg14/      # 120K embeddings, 1536-dim
│   ├── embeddings_bioclip2_vitl14/    # 120K embeddings, 768-dim
│   ├── embeddings_clip_vitl14/        # 120K embeddings, 768-dim
│   └── embeddings_siglip_vitb16/      # 120K embeddings, 768-dim
├── extreme_uneven_embeddings/     # Full dataset embeddings (PKL)
│   ├── aves_full_dinov3_embeddings.pkl      # 74,396 embeddings
│   └── mammalia_full_dinov3_embeddings.pkl  # 65,484 embeddings
└── execution_logs/                # Experiment execution logs

Quick Start

Load Primary Benchmark Results

from datasets import load_dataset

# Load main benchmark results (27,600 configurations)
ds = load_dataset("AI-EcoNet/HUGO-Bench-Paper-Reproducibility", "primary_benchmarking")
print(f"Configurations: {len(ds['train'])}")

Load Pre-computed Embeddings

# Load DINOv3 embeddings (120,000 images)
embeddings = load_dataset(
    "AI-EcoNet/HUGO-Bench-Paper-Reproducibility", 
    "embeddings_dinov3_vith16plus"
)
print(f"Embeddings shape: {len(embeddings['train'])} x {len(embeddings['train'][0]['embedding'])}")

Load Specific Analysis Results

# Model comparison results
model_comp = load_dataset("AI-EcoNet/HUGO-Bench-Paper-Reproducibility", "model_comparison")

# Scaling test results
scaling = load_dataset("AI-EcoNet/HUGO-Bench-Paper-Reproducibility", "scaling_tests")

# Intra-species variation analysis
intra = load_dataset("AI-EcoNet/HUGO-Bench-Paper-Reproducibility", "intra_species_variation")

Load Cluster Image Mappings

The intra-species analysis includes a mapping file showing which images belong to which clusters:

from huggingface_hub import hf_hub_download
import json

# Download mapping file
mapping_file = hf_hub_download(
    "AI-EcoNet/HUGO-Bench-Paper-Reproducibility",
    "intra_species_variation/cluster_image_mappings.json",
    repo_type="dataset"
)

with open(mapping_file) as f:
    mappings = json.load(f)

# Structure: {species: {run: {cluster: [image_names]}}}
print(f"Species analyzed: {list(mappings.keys())}")

Load Full Dataset Embeddings

For the extreme uneven distribution experiments, we provide full dataset embeddings:

from huggingface_hub import hf_hub_download
import pickle

# Download Aves embeddings (74,396 images)
pkl_file = hf_hub_download(
    "AI-EcoNet/HUGO-Bench-Paper-Reproducibility",
    "extreme_uneven_embeddings/aves_full_dinov3_embeddings.pkl",
    repo_type="dataset"
)

with open(pkl_file, 'rb') as f:
    data = pickle.load(f)

print(f"Embeddings: {data['embeddings'].shape}")  # (74396, 1280)
print(f"Labels: {len(data['labels'])}")
print(f"Paths: {len(data['paths'])}")

Experimental Setup

Models Evaluated

Model	Architecture	Embedding Dim	Pre-training
DINOv3	ViT-H/16+	1280	Self-supervised
DINOv2	ViT-G/14	1536	Self-supervised
BioCLIP 2	ViT-L/14	768	Biology domain
CLIP	ViT-L/14	768	Contrastive
SigLIP	ViT-B/16	768	Sigmoid loss

Clustering Methods

• K-Means, DBSCAN, HDBSCAN, Agglomerative, Spectral
• GMM (Gaussian Mixture Models)
• With and without dimensionality reduction (UMAP, t-SNE, PCA)

Evaluation Metrics

• Supervised: Adjusted Rand Index (ARI), Normalized Mutual Information (NMI), Accuracy, F1
• Unsupervised: Silhouette Score, Calinski-Harabasz Index, Davies-Bouldin Index

Citation

If you use this dataset, please cite:

@article{markoff2026vision,
  title={Vision Transformers for Zero-Shot Clustering of Animal Images: A Comparative Benchmarking Study},
  author={Markoff, Hugo and Bengtson, Stefan Hein and Ørsted, Michael},
  journal={[Journal/Conference]},
  year={2026}
}

License

This dataset is released under the CC-BY-4.0 License.

Contact

For questions or issues, please open an issue in this repository or contact the authors.