This dataset contains field camera-trap imagery from wildlife study sites. The data are licensed CC BY-NC 4.0 (non-commercial). By requesting access you agree not to redistribute the imagery, not to attempt to identify or publish the physical locations of the cameras, and to cite the paper in any work that uses it.

Camera-trap flag survey for Reference-Conditioned Distance Intervals

The flag survey behind the paper Reference-Conditioned Distance Intervals on Unseen Camera Traps (Computer Vision for Ecology Workshop, ECCV 2026). A field team placed numbered flags at known distances in front of each trail camera and photographed them with the camera itself. Each photo therefore contains pixels whose ground distance from the camera is known, which is the metric ground truth used to calibrate the cameras and to train and evaluate the model.

Code: https://github.com/toqitahamid/camtrap-distance
Model weights: https://huggingface.co/toqi/camtrap-distance


cameras	62 (Browning BTC-7E), at two forest sites in Illinois, USA (`MAS`, `TON`)
flag photos	122; 60 cameras have two photos taken months apart, 2 have one
image size	1920 × 1080
ground-contact markers after quality control	4,395
flag distances	2 to 15 m, integer metres, along three transects (left, centre, right)

flaglabel-dataset/<CAMERA>/<IMAGE>.JPG     flag photo
flaglabel-dataset/<CAMERA>/<IMAGE>.json    annotations for that photo
calibrations/<CAMERA>/<IMAGE>.json         fitted camera parameters for that photo

Annotations

Each annotation file lists, for every visible flag, its transect, its known distance, and one or more of:

the ground-contact point, where the flag meets the ground;
a span along the flag body (known physical length);
a span from the flag top to the ground.

Flags are identical, so a point fixes a pixel at a known distance and a span fixes a pixel length at a known size. Flag dimensions are stored in each file.

Calibrations

calibrations/ is derived from the annotations by scripts/calibrate.py in the code repository: focal length, camera height, pitch and roll per photo, plus quality-control diagnostics. It is included so the paper's numbers can be reproduced without refitting.

Splits

The paper's split of 43 training, 7 validation and 12 test cameras is defined in the code repository (src/network/build_targets.py); the dataset itself is not split.

Privacy

All EXIF metadata has been removed from the photos, and the originals carried no GPS. The camera banner shows only temperature, date, time and a camera code.

Usage

pip install huggingface_hub
hf auth login                              # after access is granted

From the code repository, PYTHONPATH=. python scripts/fetch_data.py downloads the survey into data/, where the calibrator, the training code and the tests expect it.

Licence

CC BY-NC 4.0 (see LICENSE), with the access conditions stated above.

Citation

@inproceedings{sarker2026reference,
  title     = {Reference-Conditioned Distance Intervals on Unseen Camera Traps},
  author    = {Sarker, Toqi Tahamid and Islam, Taminul and Morelock, Seth J.
               and Bastille-Rousseau, Guillaume and Ahmed, Khaled R.},
  booktitle = {Computer Vision for Ecology Workshop, European Conference on Computer Vision (ECCV)},
  year      = {2026}
}

Downloads last month: 12

Total file size:

79.8 MB

You need to agree to share your contact information to access this dataset