🦌 BAMBI Dataset

July 14, 2026 · View on GitHub

header image bambi dataset

🦌 BAMBI Dataset

Multimodal Nadir UAV-Recordings of Forest Wildlife

The BAMBI dataset is a large-scale airborne multispectral wildlife dataset comprising 389 paired RGB and thermal aerial video sequences recorded across diverse forest and forest-adjacent habitats in Austria. Each frame is geo-referenced with precise global coordinates (longitude, latitude, and altitude), enabling learning and evaluation in both image space and geographic space.

This repository provides sample scripts for downloading, processing, and visualizing the dataset.

Citation: If you use this dataset in your research, please cite our paper (see Citation).

If you want to make your first steps please refer to our Jupyter Notebook.

Dataset Overview

Property	Value
Video sequences	389 paired RGB + thermal
Total flight time	~45 hours
Annotated tracks	5,100
Key frames	92,701
Interpolated bounding boxes	1,218,903
Species classes	12
Recording period	January 2023 – November 2024
Location	Austria (Tyrol, Upper Austria, Lower Austria, Salzburg, Styria, Carinthia)

Species

The dataset covers the following 12 classes:

#	Species	Common Name	Wikidata	Tracks	Key Frames
1	Sus scrofa	Wild boar	Q58697	1,770	26,132
2	Cervus elaphus	Red deer	Q1219579	1,613	26,674
3	Capreolus capreolus	Roe deer	Q122069	682	9,404
4	Dama dama	Fallow deer	Q20908334	297	20,003
5	Capra ibex	Alpine ibex	Q168327	100	3,005
6	Rupicapra rupicapra	Chamois	Q131340	15	747
7	Aves	Bird	Q5113	75	942
8	Homo sapiens	Human	Q15978631	93	1,158
9	Canis lupus familiaris	Dog	Q26972265	7	95
10	Sus scrofa × Sus domesticus	Hybrid pig	Q602666 (no matching wikidata id, so workaround with forest hog)	44	1,484
11	—	No-animal	Q10738 (it is not an animal, it is a rock, THE ROCK)	60	521
12	—	Unknown	Q24238356	344	2,536

Annotation Format (MOT)

Annotations are stored as CSV files (no header) following a custom MOT format:

frame, track_id, bb_left, bb_top, bb_width, bb_height, conf, class_id, visibility, species, gender, age, is_propagated

Column	Type	Description
`frame`	int	Frame index
`track_id`	int	Unique track identifier
`bb_left`	int	Bounding box left coordinate (px)
`bb_top`	int	Bounding box top coordinate (px)
`bb_width`	int	Bounding box width (px)
`bb_height`	int	Bounding box height (px)
`conf`	float	Confidence score
`class_id`	int	Species class identifier
`visibility`	float	Visibility (1.0 = fully visible, 0.0 = fully occluded)
`species`	str	Species name
`gender`	int	Gender (0 = unknown, 1 = male, 2 = female)
`age`	int	Age (0 = unknown, 1 = juvenile, 2 = adult)
`is_propagated`	int	0 = annotated key frame, 1 = interpolated

Class_ids are defined as <wikidata_id>-<gender>-<age>-<visibility> with the following mapping:

{
  "class_mapping": {
    "Q10738-0-0-0": 0,
    "Q10738-0-0-1": 1,
    "Q1219579-0-0-0": 2,
    "Q1219579-0-0-1": 3,
    "Q1219579-0-1-0": 4,
    "Q1219579-0-1-1": 5,
    "Q1219579-0-2-0": 6,
    "Q1219579-0-2-1": 7,
    "Q1219579-1-0-0": 8,
    "Q1219579-1-0-1": 9,
    "Q1219579-1-2-0": 10,
    "Q1219579-1-2-1": 11,
    "Q1219579-2-0-0": 12,
    "Q1219579-2-0-1": 13,
    "Q1219579-2-1-0": 14,
    "Q1219579-2-2-0": 15,
    "Q1219579-2-2-1": 16,
    "Q122069-0-0-0": 17,
    "Q122069-0-0-1": 18,
    "Q122069-0-1-0": 19,
    "Q122069-0-1-1": 20,
    "Q122069-0-2-0": 21,
    "Q122069-0-2-1": 22,
    "Q131340-0-0-0": 23,
    "Q131340-0-0-1": 24,
    "Q131340-0-2-0": 25,
    "Q131340-0-2-1": 26,
    "Q15978631-0-0-0": 27,
    "Q15978631-0-0-1": 28,
    "Q15978631-0-2-0": 29,
    "Q15978631-0-2-1": 30,
    "Q15978631-2-0-0": 31,
    "Q15978631-2-2-0": 32,
    "Q15978631-2-2-1": 33,
    "Q168327-0-0-0": 34,
    "Q168327-0-0-1": 35,
    "Q168327-0-1-0": 36,
    "Q168327-0-1-1": 37,
    "Q168327-0-2-0": 38,
    "Q168327-0-2-1": 39,
    "Q168327-2-2-0": 40,
    "Q168327-2-2-1": 41,
    "Q20908334-0-0-0": 42,
    "Q20908334-0-1-0": 43,
    "Q20908334-1-0-0": 44,
    "Q20908334-1-1-0": 45,
    "Q20908334-1-2-0": 46,
    "Q20908334-2-2-0": 47,
    "Q20908334-2-2-1": 48,
    "Q24238356-0-0-0": 49,
    "Q24238356-0-0-1": 50,
    "Q24238356-0-2-0": 51,
    "Q24238356-0-2-1": 52,
    "Q26972265-0-0-0": 53,
    "Q26972265-0-2-0": 54,
    "Q26972265-0-2-1": 55,
    "Q5113-0-0-0": 56,
    "Q5113-0-0-1": 57,
    "Q5113-0-2-0": 58,
    "Q5113-0-2-1": 59,
    "Q58697-0-0-0": 60,
    "Q58697-0-0-1": 61,
    "Q58697-0-1-0": 62,
    "Q58697-0-1-1": 63,
    "Q58697-0-2-0": 64,
    "Q58697-0-2-1": 65,
    "Q58697-1-2-0": 66,
    "Q58697-1-2-1": 67,
    "Q58697-2-2-0": 68,
    "Q58697-2-2-1": 69,
    "Q602666-0-0-0": 70,
    "Q602666-0-0-1": 71,
    "Q602666-0-1-0": 72,
    "Q602666-0-2-0": 73,
    "Q602666-0-2-1": 74
  }
}

Availability

The processed dataset is publicly available on Zenodo:

The unprocessed dataset is also publicly available on Zenodo:

The matched Red Deer subset:

Part 1: 10.5281/zenodo.21291168
Part 2: 10.5281/zenodo.21298044

The matched, orthographic Red Deer subset:

Scripts

Installation

pip install -r requirements.txt

The scripts are tested with Python 3.10+.

Automatic Download

Selectively download flight ZIPs from the BAMBI dataset hosted on Zenodo. Uses the zenodo_upload_summary_*.json files to resolve which depositions contain which flights, so you can grab exactly what you need without fetching entire multi-GB depositions. Use filter_flights.py to get a list of flight IDs for the data that you are looking for (e.g. filtered for species).

# List all available flights
python download_from_zenodo.py --list

# Download flights 0, 5, and 12
python download_from_zenodo.py -f 0 5 12

# Download flights 10 through 25
python download_from_zenodo.py --range 10 25

# Download all flights from parts 1 and 3
python download_from_zenodo.py --parts 1 3

# Download all flights of a dataset split (train / val / test)
python download_from_zenodo.py --split train
python download_from_zenodo.py --split val
python download_from_zenodo.py --split test

# Download everything, extract, and clean up ZIPs
python download_from_zenodo.py --unzip

# Preview what a full download would do
python download_from_zenodo.py --dry-run

# Download a different dataset version instead of the pre-processed videos (compatible with all other flags like -f, --range, --split, etc.)
python download_from_zenodo.py --version raw
python download_from_zenodo.py --version matched
python download_from_zenodo.py --version orthographic

Note: --version selects which dataset version to download and defaults to base (the pre-processed videos). The available versions are base, raw, matched, and orthographic; each is described by its own flight_metadata/zenodo_upload_summary_*.json. A summary file from a custom location can be supplied with -s <path>, which overrides --version.

Note: --split reads flight IDs from flight_metadata/splits.json. A custom path can be supplied with --splits-file <path>. The flag is silently ignored when -f, --range, or --parts is also specified.

Flight filter

Filter flights based on metadata JSON files by species, occlusion, sex, age, weather, date range, and drone name.

All filters combine with AND logic between each other. Within list filters (--species, --drone, --sex, --age) values combine with OR logic. Weather flags combine with AND (all specified conditions must be present).

# Multiple species (OR: flights containing either)
python filter_flights.py --species "Roe deer" "Homo sapiens" "Q122069"

# Only flights with occluded frames
python filter_flights.py --occlusion true

# Flights with male or female subjects
python filter_flights.py --sex male female

# Flights with juvenile or adult animals
python filter_flights.py --age juvenile adult

# Flights that are both cloudy AND windy
python filter_flights.py --weather cloudy windy

# Flights in October 2024
python filter_flights.py --min-date 2024-10-01 --max-date 2024-10-31

# Visible roe deer in sunny weather during October 2024
python filter_flights.py ./metadata \
    --species "Roe deer" \
    --occlusion false \
    --weather sunny \
    --min-date 2024-10-01 \
    --max-date 2024-10-31 \
    -v

Frame Extraction

Extract thermal and RGB frames from the side-by-side video files. The left half of each frame contains the thermal channel, the right half the RGB channel.

# Extract every frame
python frame_extraction.py video.mp4 ./frames

# Extract every 10th frame, RGB only
python frame_extraction.py video.mp4 ./frames --sample-rate 10 --thermal false

# Extract a specific frame range
python frame_extraction.py video.mp4 ./frames --start 1000 --end 2000

Output structure:

frames/
├── thermal/
│   ├── video_00001000.png
│   ├── video_00001001.png
│   └── ...
└── rgb/
    ├── video_00001000.png
    ├── video_00001001.png
    └── ...

MOT Interpolation

Interpolate missing frames between annotated key frames using linear interpolation of bounding box coordinates. Interpolated entries are marked with is_propagated=1.

# Interpolate a single annotation file
python mot_interpolation.py annotations.txt ./output

# Interpolate all files in a folder with a custom frame step
python mot_interpolation.py ./annotations_folder ./output --step 2

MOT Filter

Filter annotation files by species, class ID, visibility, bounding box size, gender, or age.

# Keep only wild boar annotations
python mot_filter.py annotations.txt -o ./filtered --species "Sus scrofa (Wild boar)"

# Filter by class ID and minimum bounding box width
python mot_filter.py ./annotations/ -o ./filtered --class-id 50 51 --min-width 25

# Combine multiple filters
python mot_filter.py annotations.txt -o ./filtered --species "Cervus elaphus (Red deer)" --visibility 1.0 --min-width 30

MOT to YOLO Conversion

Convert MOT annotations to YOLO label format for training object detection models.

# Convert with default label (class_id only)
python mot_to_yolo.py annotations.txt -o ./yolo_labels

# Include species, gender, and age in the label
python mot_to_yolo.py ./annotations/ -o ./yolo_labels --img-width 640 --img-height 512 --labels class_id gender age

Visualization

On Extracted Frames

Overlay bounding boxes on individual extracted frames:

# Visualize key frame annotations
python mot_frame_viewer.py frame_image.png annotations.txt --show

# With interpolation of in-between frames
python mot_frame_viewer.py frame_image.png annotations.txt --interpolate --show

# Save the visualization to a file
python mot_frame_viewer.py frame_image.png annotations.txt -o output.png

On Video

Overlay bounding box tracks directly on video:

# Live preview
python mot_video_viewer.py video.mp4 annotations.txt --show

# Save annotated video
python mot_video_viewer.py video.mp4 annotations.txt -o annotated_output.mp4

# With interpolated tracks
python mot_video_viewer.py video.mp4 annotations.txt -o output.mp4 --interpolate

DEM from Poses

Download, merge, and clip Digital Elevation Model (DGM) data from the Austrian Federal Office of Metrology and Surveying (BEV) based on GPS coordinates from pose files. Outputs a GeoTIFF clipped to the flight area, with optional GLB mesh generation.

# From a single poses file
python dem_from_poses.py --file recordings/0_matched_poses.json

# From a folder (finds all *_poses.json recursively)
python dem_from_poses.py --folder recordings/

# Custom padding and output directory
python dem_from_poses.py --folder recordings/ --padding 100 --output-dir /tmp/dems

# Custom simplification factor and output CRS
python dem_from_poses.py --file recording/0_matched_poses.json --simplify 1 --output-crs EPSG:32633

# Skip mesh generation
python dem_from_poses.py --file recording/0_matched_poses.json --no-mesh

Option	Default	Description
`--file` / `--folder`	(required)	Single `*_poses.json` file or folder to scan recursively
`--padding`	`50`	Padding around bounding box in metres
`--output-dir`	`DEM/` subfolder	Output directory for GeoTIFF and mesh files
`--output-crs`	`EPSG:32633`	Output coordinate reference system
`--cache-dir`	`~/.cache/austria_dem`	Tile cache directory
`--force-download`	`false`	Re-download tiles even if cached
`--no-mesh`	`false`	Skip GLTF (`.glb`) mesh generation
`--simplify`	`2`	Mesh simplification factor

Add Relative DEM Position to Poses

Add relative location offsets to drone pose files based on DEM origin metadata created with dem_from_poses.py. Converts WGS84 (lat/lng/alt) coordinates to the CRS defined in the DEM metadata, then computes relative [x, y, z] offsets from the DEM origin. Also adds a rotation field ([pitch, roll, yaw]) to each image entry.

# Single file pair
python add_relative_dem_position_to_poses.py --poses 0_matched_poses.json --dem 0_matched_dem.json --output ./output

# Folder mode (matches files by filename prefix)
python add_relative_dem_position_to_poses.py --poses ./poses/ --dem ./dem_metadata/ --output ./output

# Modify pose files in place
python add_relative_dem_position_to_poses.py --poses ./poses/ --dem ./dem_metadata/ --inplace

Option	Default	Description
`--poses`	(required)	Single pose JSON file or folder of `*_matched_poses.json` files
`--dem`	(required)	Single DEM metadata JSON file or folder of `*_matched_dem.json` files
`--output`	—	Output folder for modified pose files
`--inplace`	`false`	Modify pose files in place (requires no `--output`)

In folder mode, files are matched by the prefix before the first underscore (e.g., 0_matched_poses.json matches 0_matched_dem.json).

RGB–Thermal Frame Matching

Although the drones record RGB and thermal imagery simultaneously, the two modalities are not temporally synchronized. Temporal offsets vary with flight dynamics, so annotations from thermal frames cannot be directly transferred to RGB images — correspondence must be established at sequence or frame level.

The dataset includes a local patch-based matching strategy to align thermal annotations with their RGB counterparts:

For each thermal detection, the annotated crop is extracted.
A larger search region is generated in the corresponding RGB frame by expanding the bounding box.
The RGB region is converted to grayscale and processed using multiple transformations (e.g., CLAHE and edge-based filtering), producing five different variants.
Each variant is matched to the thermal crop via a sliding-window approach, selecting the position with the highest matching confidence.
For frames with multiple detections, alignment is accepted if pixel shift estimates across methods agree within ±10 pixels, and the most consistent candidate is selected.
For frames with a single detection, a stricter majority consensus is required; otherwise the sample is discarded.

The implementation is available in a separate repository:

🔗 BAMBI BBox Corrections

AlfsPY: Framework for orthographic projections and light field renderings based on the drone recordings.
Detection: Examples on using AlfsPY for different tasks like geo-tiff generation.
Geo-Referenced Tracking: Implementation of tracking algorithms based on local image as well as global world coordinates.
Bambi-QGIS: Plugin for integrating drone video processing to the geo-information system QIGS.

Citation

@misc{praschl2026bambi,
  title        = {The BAMBI Dataset: Multimodal Nadir UAV-Recordings of Forest Wildlife},
  author       = {Praschl, Christoph and Markoff, Hugo and Maschek, Anna and Jantsch, Wolfram and Wohlfahrt, Stephanie and Leitner, Horst and Beery, Sara and {\O}rsted, Michael and Schedl, David C.},
  year         = {2026},
  howpublished = {Presented at the CV4Animals Workshop, held in conjunction with the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
  note         = {Non-archival workshop paper},
}

License

This repository is licensed under the MIT License.