Replication Dataset: Oh Deer, What Should I Do With This? Seasonal Priors for Selective Wildlife Annotation and Classification

This dataset is part of the BAMBI project (Biodiversity Airborne Monitoring Based on Intelligent UAV Sampling; https://www.bambi.eco/). It provides a multimodal RGB–thermal crop dataset for fine-grained red deer annotation and classification from nadir UAV imagery.

The dataset contains 7,295 matched RGB–thermal crop pairs of red deer instances recorded in Austrian forest environments. Each sample consists of one RGB crop and one corresponding thermal crop showing the same detected animal instance. The crops are derived from synchronized UAV recordings and are intended for studying fine-grained wildlife classification under challenging aerial imaging conditions, where animals occupy only small image regions and relevant visual cues may be weak, occluded, or seasonally absent.

The dataset focuses on adult-male identification in red deer. Since antlers are the main visual cue for distinguishing male red deer, and juveniles do not have antlers either, the task is formulated as a binary fine-grained classification problem: adult male versus female/juvenile. Ambiguous cases are not forced into the binary label space. Instead, the annotation protocol explicitly includes uncertain and occluded as abstention states.

Each crop pair was annotated independently by three annotators with a wildlife ecology background under multiple viewing conditions:

• RGB-only

• thermal-only

• matched RGB+thermal

For each condition, annotators assigned one of four labels: male, female_juvenile, uncertain, or occluded. The matched RGB+thermal annotations are intended to provide the primary multimodal reference labels, while the RGB-only and thermal-only annotations allow modality-specific analyses of label uncertainty, disagreement, and evidence sufficiency.

In addition to the full annotation files, the dataset contains reduced annotation subsets for samples located in modality-specific uncertainty periods based on the seasonal red deer antler cycle. These uncertainty-focused annotation files support experiments on biologically grounded seasonal priors, selective annotation, and abstention-aware classification.

The dataset is organized as follows:

Dataset/
├── rgb_crops_matched/
│   └── 7,295 RGB crop images
├── thermal_crops_matched/
│   └── 7,295 corresponding thermal crop images
├── annotations/
│   ├── annotator_1/
│   ├── annotator_2/
│   └── annotator_3/
│       ├── matched_all.json
│       ├── matched_uncertainty.json
│       ├── rgb_all.json
│       ├── rgb_uncertainty.json
│       ├── thermal_all.json
│       └── thermal_uncertainty.json
├── metadata_flights/
│   └── per-flight metadata files
└── configs/
    ├── splits/
    │   ├── common_pool_grouped_5fold.json
    │   ├── common_pool_grouped_5fold_flights.csv
    │   ├── shared_seed42_triplet_yolo_split.json
    │   └── shared_seed42_triplet_yolo_split_flights.csv
    └── tests/
        ├── modality_band_testing.json
        ├── model_family_comparison.json
        ├── prior_weighting.json
        ├── channel_replacement.json
        └── paper_tests_manifest.json

The crop files follow the naming convention:

<flight_id>_<frame_idx>_<instance_idx>.jpg

where flight_id identifies the original UAV flight, frame_idx identifies the source video frame, and instance_idx distinguishes multiple animal instances within the same frame. Matching RGB and thermal crops share the same filename and are stored in the respective modality folders.

The metadata_flights/ folder contains per-flight metadata, including flight identifiers, recording dates, drone information, frame counts, split assignments, and species information. The dataset covers 56 flights across nine calendar months with sufficient data coverage: January, February, March, April, May, July, September, November, and December.

The configs/splits/ folder provides the exact grouped flight-level splits used for the associated experiments. The grouped five-fold split is intended for common matched-label evaluation without flight leakage, while the shared seed-42 split is provided for model-family comparisons involving triplet-based and YOLO-based classifiers.

The configs/tests/ folder contains experiment configuration files for reproducing the paper-scoped evaluations, including modality band testing, model-family comparison, prior weighting with seasonal antler priors, and channel-replacement experiments for RGB–thermal fusion.

The dataset is designed as a benchmark for evaluating when RGB, thermal, or matched RGB+thermal information is sufficient for fine-grained wildlife classification. It can also be used to study how biological seasonality affects both human annotation reliability and model confidence.

The crop images are derived from the original BAMBI UAV wildlife dataset. The original full-video dataset is split across multiple Zenodo records and can be found here:

• 10.5281/zenodo.18692354
• 10.5281/zenodo.18698508
• 10.5281/zenodo.18703312
• 10.5281/zenodo.18705705
• 10.5281/zenodo.18707610
• 10.5281/zenodo.18711217
• 10.5281/zenodo.18715162
• 10.5281/zenodo.18717601

Scripts for working with the original BAMBI dataset, including downloading the corresponding flight videos and extracting specific frames, are available at:

https://github.com/bambi-eco/Dataset