Depth Estimation — Metric Distance & Height for Wildlife Camera Traps

Official implementation accompanying:

A robust metric distance and height estimation pipeline for wildlife camera trap imagery Muhammad Aamir, Matthew Wijers, Andrew Loveridge, Andrew Markham Ecological Informatics, Volume 92, 103520 (2025). https://doi.org/10.1016/j.ecoinf.2025.103520

The pipeline turns raw camera-trap photos into metric camera-to-animal distances and heights, without requiring a laser rangefinder, a calibrated rig, or reliable EXIF/metadata. It combines animal detection, instance segmentation and monocular depth estimation (MDE) with an automated, reference-object-based calibration step that converts relative depth/disparity into real-world metres. It extends the original distance-sampling methodology of Haucke et al. (2022) with several newer MDE backbones, more robust calibration regression methods, updated animal detectors, and height estimation in addition to distance.

Both a desktop GUI and a command-line interface are provided for Windows, macOS (Intel & Apple Silicon) and Linux.

Table of Contents

• Overview
• Supported Models
• Installation
• Usage
  • Graphical Interface
  • Command Line
• Data Structure
• Example Datasets
• Output
• Building From Source
• Citation
• Acknowledgements
• License

Overview

For each transect (camera-trap station), the pipeline:

1. Calibrates the depth model using photos of reference objects placed at known distances (calibration_frames), fitting a monotonic curve that maps the model's relative disparity output to metric depth.
2. Detects animals (and optionally people/vehicles) in each photo using MegaDetector.
3. Segments each detection with Segment Anything (SAM) to isolate the animal from background/occluding vegetation.
4. Estimates depth for the photo using the selected monocular depth estimation model, applying the fitted calibration where the model only produces relative depth.
5. Samples a representative depth value for each detection (bounding-box percentile, bounding-box bottom, or SAM-mask interior) and converts it, together with the detection's pixel position and the camera's field of view, into a metric 3D world position — yielding both radial distance and height of the animal relative to the camera.

Results are written per-transect to CSV/TXT files suitable for downstream distance-sampling and density-estimation workflows, with optional diagnostic PDF figures for calibration and per-detection sampling.

Supported Models

Monocular Depth Estimation (--depth_estimation_model)

_{* the enum name preserves a historical typo (ahything) from the codebase.}

Metric models (Depth Anything metric, Depth Pro) bypass reference-object calibration entirely and require sample_from=detection. All other models are relative/disparity models and are calibrated automatically per transect from the calibration_frames.

The repository also contains standalone wrappers for UniDepth v2 (unidepthv2.py), an ONNX build of Depth Pro (depth_pro.py) and MiDaS v2.1 (MiDas.py); these are not currently exposed through the --depth_estimation_model option but can be wired in following the pattern of the other model files.

Animal Detection (--detection_model)

Detections can alternatively be supplied as a precomputed CSV (--box_data_csv, e.g. exported from TrapTagger), skipping detection entirely.

Segmentation

Segment Anything (SAM, ViT-B) is used to obtain a pixel-accurate instance mask for each detection when --detection_sampling_method sam is selected.

Calibration Regression (--calibration_regression_method)

ransac (default) · leastsquares · poly · ransac_poly · piecewise_linear — fit to the relationship between each calibration frame's disparity and a reference frame, in either disparity or metric-depth space (--calibrate_metric).

Installation

git clone https://github.com/Morphocam/Depth-Estimation.git
cd Depth-Estimation

# Python 3.10 or 3.11 recommended (3.12 has known GTK-binding build issues on Linux)
conda create --name depth-estimation python=3.10
conda activate depth-estimation

pip install -r requirements.txt

Pre-built executables for Windows, macOS and Linux are also published under Releases — no Python installation required.

Model Weights

The ONNX-based models (DPT, MegaDetector v5a, SAM, Depth Anything, Metric3D v2) are downloaded automatically on first use, or can be fetched in advance:

# Linux/macOS
./weights/download.sh

# Windows (PowerShell)
./weights/download.ps1

MegaDetector v5b requires manually placing md_v5b.0.0.pt inside weights/ (not auto-downloaded due to file size). MegaDetector v6, Depth Pro, MiDaS v3.1 and Monodepth2 weights are fetched automatically via PytorchWildlife, transformers/HuggingFace Hub, torch.hub, and the official Monodepth2 release respectively, the first time each model is selected.

Usage

Graphical Interface

python main.py

Select your data directory (see Data Structure), configure the detection/depth/calibration/sampling options in the scrollable panel, and press Start. Configuration is persisted between runs.

Command Line

python main.py --cli --help

Every configuration field in config.py is automatically exposed as a CLI flag (booleans become --flag/--no_flag, enums accept lowercase names, e.g. --depth_estimation_model dpt_pytorch). Typical usage:

python main.py --cli \
  --data_dir /path/to/DepthEstimationData \
  --depth_estimation_model dpt \
  --detection_model megadetector_v5a \
  --calibration_regression_method ransac \
  --detection_sampling_method bbox_percentile \
  --make_figures

Data Structure

The data directory is user-named and must contain a results directory and a transects directory, with one subdirectory per camera-trap station:

DepthEstimationData
├── results
└── transects
    ├── transect_001
    │   ├── calibration_frames
    │   │   ├── 5.png
    │   │   ├── 10.png
    │   │   └── ...
    │   ├── calibration_frames_masks
    │   │   ├── 5.png
    │   │   ├── 10.png
    │   │   └── ...
    │   └── detection_frames
    │       ├── detection_000001.png
    │       └── ...
    └── ...

• calibration_frames — photos of reference objects, each file named after the distance in metres it represents (e.g. 10.png, or 10m), or accompanied by a .txt file containing the distance.
• calibration_frames_masks — one equally-named binary mask per calibration photo, with the reference object marked white and everything else black.
• detection_frames — arbitrarily named photos of the animals to estimate distance/height for. .jp(e)g and .png are supported.

Optional extensions:

• Precomputed metric depth — placing a detection_frames_depth directory alongside detection_frames (matching filenames, .exr by default) lets a transect supply ready-made depth maps and skip the depth model entirely for that image.
• Precomputed detections — --box_data_csv path/to/boxes.csv supplies bounding boxes (columns Site.Name, Image.ID, Score, Left, Right, Top, Bottom) instead of running a detector, matched against detection-frame filenames.
• Legacy _cropped directories — calibration_frames_cropped/detection_frames_cropped are still accepted for backwards compatibility; new datasets should use the --crop_top/bottom/left/right flags instead.

Example Datasets

• datasets/Oxford_station — controlled ground-truth data collected on a surveyed grid at the WildCRU research base, Oxfordshire, UK.
• datasets/BVC_stations — real-world camera-trap footage from Bubye Valley Conservancy, Zimbabwe, used to validate the pipeline in the field.

Both are laid out in the Data Structure format above and are the field/ground-truth datasets used in the accompanying paper.

Output

For every processed transect, results are written to the results directory:

• results.csv — one row per detection: transect_id, frame_id, detection_idx, detection_confidence, depth, world_x, world_y, world_z, error_status. world_y corresponds to the estimated height, world_z/radial position to distance.
• results.txt — a simplified Camera trap / Label / Observation / Radial distance table for direct use with downstream distance-sampling and abundance-estimation tools.
• calibration/*.pdf and sampling/*.pdf (only with --make_figures) — diagnostic depth-map and detection visualizations, optionally annotated with detection IDs (--draw_detection_ids) and estimated world positions (--draw_world_position).

Building From Source

Standalone executables are produced with PyInstaller via main.spec:

pyinstaller main.spec

• Linux — Dockerfile.ubuntu22 builds a Ubuntu 22.04-compatible executable, including GTK bindings needed by the Toga GUI on Linux.
• macOS — after building the .app bundle, dmgbuild_settings.py packages it into a .dmg installer; scripts/macos_codesign.sh handles signing and notarization for releases.
• Windows — main.spec produces a standalone .exe directly.

Citation

If you use this software or build on this work, please cite:

@article{aamir2025robust,
  title   = {A robust metric distance and height estimation pipeline for wildlife camera trap imagery},
  author  = {Aamir, Muhammad and Wijers, Matthew and Loveridge, Andrew and Markham, Andrew},
  journal = {Ecological Informatics},
  volume  = {92},
  pages   = {103520},
  year    = {2025},
  issn    = {1574-9541},
  doi     = {10.1016/j.ecoinf.2025.103520},
  url     = {https://www.sciencedirect.com/science/article/pii/S1574954125005291}
}

This pipeline builds on the original distance-sampling methodology, please also consider citing:

@article{haucke2022overcoming,
  title   = {Overcoming the distance estimation bottleneck in estimating animal abundance with camera traps},
  author  = {Haucke, Timm and K{\"u}hl, Hjalmar S. and Hoyer, Jacqueline and Steinhage, Volker},
  doi     = {https://doi.org/10.1016/j.ecoinf.2021.101536},
  journal = {Ecological Informatics},
  volume  = {68},
  pages   = {101536},
  year    = {2022},
  issn    = {1574-9541},
  url     = {https://www.sciencedirect.com/science/article/pii/S1574954121003277}
}

The pipeline additionally builds on the following third-party models — please cite the ones relevant to your configuration:

@article{megadetector,
  title   = {Efficient pipeline for camera trap image review},
  author  = {Beery, Sara and Morris, Dan and Yang, Siyu},
  journal = {arXiv preprint arXiv:1907.06772},
  year    = {2019}
}
@inproceedings{dpt,
  title     = {Vision transformers for dense prediction},
  author    = {Ranftl, Ren{\'e} and Bochkovskiy, Alexey and Koltun, Vladlen},
  booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision},
  pages     = {12179--12188},
  year      = {2021}
}
@article{midas3.1,
  title   = {MiDaS v3.1 -- A Model Zoo for Robust Monocular Relative Depth Estimation},
  author  = {Birkl, Reiner and Wofk, Diana and M{\"u}ller, Matthias},
  journal = {arXiv preprint arXiv:2307.14460},
  year    = {2023}
}
@article{segment_anything,
  title   = {Segment Anything},
  author  = {Kirillov, Alexander and Mintun, Eric and Ravi, Nikhila and Mao, Hanzi and Rolland, Chloe and Gustafson, Laura and Xiao, Tete and Whitehead, Spencer and Berg, Alexander C. and Lo, Wan-Yen and Doll{\'a}r, Piotr and Girshick, Ross},
  journal = {arXiv:2304.02643},
  year    = {2023}
}
@inproceedings{depthanything,
  title     = {Depth Anything: Unleashing the Power of Large-Scale Unlabeled Data},
  author    = {Yang, Lihe and Kang, Bingyi and Huang, Zilong and Xu, Xiaogang and Feng, Jiashi and Zhao, Hengshuang},
  booktitle = {CVPR},
  year      = {2024}
}
@article{depthpro,
  title   = {Depth Pro: Sharp Monocular Metric Depth in Less Than a Second},
  author  = {Bochkovskii, Aleksei and Delaunoy, Ama{\"e}l and Germain, Hugo and Santos, Marcel and Zhou, Yichao and Richter, Stephan R. and Koltun, Vladlen},
  journal = {arXiv preprint arXiv:2410.02073},
  year    = {2024}
}
@inproceedings{unidepth,
  title     = {UniDepth: Universal Monocular Metric Depth Estimation},
  author    = {Piccinelli, Luigi and Yang, Yung-Hsu and Sakaridis, Christos and Segu, Mattia and Li, Siyuan and Van Gool, Luc and Yu, Fisher},
  booktitle = {CVPR},
  year      = {2024}
}
@article{metric3dv2,
  title   = {Metric3Dv2: A Versatile Monocular Geometric Foundation Model for Zero-shot Metric Depth and Surface Normal Estimation},
  author  = {Hu, Mu and Yin, Wei and Zhang, Chi and Cai, Zhipeng and Long, Xiaoxiao and Chen, Hao and Wang, Kaixuan and Yu, Gang and Shen, Chunhua and Shen, Shaojie},
  journal = {IEEE Transactions on Pattern Analysis and Machine Intelligence},
  year    = {2024}
}
@inproceedings{monodepth2,
  title     = {Digging into Self-Supervised Monocular Depth Estimation},
  author    = {Godard, Cl{\'e}ment and Mac Aodha, Oisin and Firman, Michael and Brostow, Gabriel J.},
  booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision},
  year      = {2019}
}
@article{pytorchwildlife,
  title   = {Pytorch-Wildlife: A Collaborative Deep Learning Framework for Conservation},
  author  = {Hernandez, Andres and Miao, Zhongqi and Vargas, Luisa and Beery, Sara and Dodhia, Rahul and Lavista, Juan},
  journal = {arXiv preprint arXiv:2405.12930},
  year    = {2024}
}

Acknowledgements

This work was supported by the Natural Environment Research Council. The authors extend their sincere gratitude to the management and research staff of the Bubye Valley Conservancy for their support and assistance with data collection. We also thank the Paul G. Allen Family Foundation for its generous funding contributions to the WildCAT Trust in Zimbabwe, which made the Bubye Valley Conservancy survey possible.

License

Released under the MIT License.