--- license: cc-by-4.0 task_categories: - image-classification - zero-shot-classification tags: - biology - ecology - wildlife - camera-traps - vision-transformers - clustering - zero-shot-learning - biodiversity - reproducibility - benchmarking - embeddings - dinov3 - dinov2 - bioclip - clip - siglip language: - en pretty_name: HUGO-Bench Paper Reproducibility Data size_categories: - 100K **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 species-level clustering of camera trap images. ### Relationship to HUGO-Bench This dataset is derived from [HUGO-Bench](https://huggingface.co/datasets/AI-EcoNet/HUGO-Bench), which provides the source images and species annotations. While HUGO-Bench contains the **validated image crops** (139,111 images across 60 species), this repository provides: - **Clustering results** from all 27,600 experimental configurations - **Pre-computed embeddings** enabling reproduction without image access - **Execution logs** for full experimental traceability | Dataset | Content | Purpose | |---------|---------|---------| | [HUGO-Bench](https://huggingface.co/datasets/AI-EcoNet/HUGO-Bench) | 139,111 validated camera trap images | Source images for experiments | | **This repository** | Results, embeddings, logs | Paper reproducibility | ## Repository Structure ``` ├── 01_primary_benchmarking/ # Full 27,600 configuration results │ ├── clustering_analysis_complete.csv │ ├── clustering_analysis_with_ami.csv │ ├── comprehensive_vmeasure_by_class.json │ └── images_run_*.json # Subsample definitions (10 runs) │ ├── 02_model_comparison/ # 5 ViT model comparison │ ├── dinov3_all_combinations_results.json │ ├── dinov3_bioclip_siglip_all_methods_results.json │ └── dinov3_comparison_results.json │ ├── 03_dimensionality_reduction/ # t-SNE, UMAP, PCA, Isomap, KPCA │ └── dimensionality_comparison.json │ ├── 04_clustering_supervised/ # K-variation experiments (K=15,30,45,90,180) │ ├── k30_metrics_by_class.json │ └── k_variation_by_dimred_class.json │ ├── 05_clustering_unsupervised/ # HDBSCAN vs DBSCAN │ └── unsupervised_metrics_by_class.json │ ├── 06_cluster_count_prediction/ # Progressive species testing (1,200 runs) │ ├── progressive_species_testing_results.json │ └── progressive_species_testing_results_expanded.json │ ├── 07_intra_species_variation/ # Age, sex, pelage detection │ ├── wolf_dbscan_clusters/ │ └── intra_cluster/ │ ├── 08_uneven_distribution/ # Long-tailed distribution tests │ ├── extreme_20_max_test/ │ ├── original_config_extreme_uneven_test/ │ └── even_distribution_results.json │ ├── 09_scaling_tests/ # 5-60 species scaling behavior │ ├── scaling_test_results/ │ └── different_n_test/ │ ├── 10_embeddings/ # Pre-computed embeddings │ ├── embeddings/ # Standard benchmarking embeddings │ ├── extreme_uneven_embeddings/ │ └── extreme_uneven_image_lists/ │ └── execution_logs/ # Complete execution logs ├── clustering_dimred_log.txt ├── clustering_complete_log.txt └── ... ``` ## Key Results Summary Our benchmarking evaluated **27,600 configurations** across: - **5 ViT Models**: DINOv3, DINOv2, BioCLIP 2, CLIP, SigLIP - **5 Dimensionality Reduction**: t-SNE, UMAP, PCA, Isomap, Kernel PCA - **4 Clustering Algorithms**: Hierarchical, GMM, HDBSCAN, DBSCAN - **60 Species**: 30 mammals + 30 birds from camera trap imagery ### Top Performing Configuration | Component | Best Choice | V-Measure | |-----------|-------------|-----------| | Model | DINOv3 | 0.958 | | Dim. Reduction | t-SNE | +26-38pp vs others | | Clustering (supervised) | Hierarchical K=30 | 0.958 | | Clustering (unsupervised) | HDBSCAN | 0.943 | ## Usage ### Loading Results with Python ```python import pandas as pd import json # Load primary benchmarking results results = pd.read_csv("01_primary_benchmarking/clustering_analysis_complete.csv") # Filter for best model dinov3_results = results[results['model'] == 'dinov3'] # Load JSON metrics with open("05_clustering_unsupervised/unsupervised_metrics_by_class.json") as f: unsupervised = json.load(f) ``` ### Using Pre-computed Embeddings The `10_embeddings/` folder contains pre-computed embeddings that allow running clustering experiments **without needing the original images**: ```python import numpy as np import json # Load embeddings embeddings = np.load("10_embeddings/embeddings/dinov3_embeddings.npy") # Load corresponding image list with open("01_primary_benchmarking/images_run_1.json") as f: image_list = json.load(f) ``` ### Reproducing Paper Tables Each folder corresponds to specific paper sections: | Paper Section | Data Folder | |--------------|-------------| | Table 3 (V-measure by model) | `01_primary_benchmarking/` | | Table 4 (Dim. reduction comparison) | `03_dimensionality_reduction/` | | Table 5 (Supervised K variation) | `04_clustering_supervised/` | | Table 6 (Unsupervised comparison) | `05_clustering_unsupervised/` | | Figure 5 (Cluster count prediction) | `06_cluster_count_prediction/` | | Table 7 (Intra-species traits) | `07_intra_species_variation/` | | Table 8 (Uneven distribution) | `08_uneven_distribution/` | | Figure 8 (Scaling behavior) | `09_scaling_tests/` | ## File Formats | Extension | Description | How to Load | |-----------|-------------|-------------| | `.csv` | Tabular results | `pandas.read_csv()` | | `.json` | Structured metrics | `json.load()` | | `.npy` | NumPy embeddings | `numpy.load()` | | `.txt`/`.log` | Execution logs | Plain text | ## Citation If you use this data, please cite both the paper and HUGO-Bench: ```bibtex @article{markoff2025vit_clustering, title={Vision Transformers for Zero-Shot Clustering of Animal Images: A Comparative Benchmarking Study}, author={Markoff, Hugo and Bengtson, Stefan Hein and {\O}rsted, Michael}, journal={TBD}, year={2025} } @dataset{hugo_bench, title={HUGO-Bench: A Benchmark Dataset for Camera Trap Species Clustering}, author={AI-EcoNet}, year={2025}, url={https://huggingface.co/datasets/AI-EcoNet/HUGO-Bench} } ``` ## License This dataset is released under [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/). ## Contact - **Hugo Markoff** - khbm@bio.aau.dk - Department of Chemistry and Bioscience, Aalborg University ## Related Resources - 📊 [HUGO-Bench Dataset](https://huggingface.co/datasets/AI-EcoNet/HUGO-Bench) - Source images (139,111 validated crops) - 💻 [GitHub Repository](https://github.com/HugoMarkoff/animal_visual_transformer) - Code and scripts - 🌐 [Interactive Visualization](https://hugomarkoff.github.io/animal_visual_transformer/) - Explore clustering results