DeepLabCut pose estimation models for wild gorilla locomotor biomechanics

This repository contains trained DeepLabCut (v3.0) models, OpenApePose outputs, and associated evaluation data for automated pose estimation of wild western lowland gorillas (Gorilla gorilla gorilla), developed as part of King et al. (Advancing methods for quantitative biomechanics in wild apes through deep learning pose estimation, in preparation).
Six DeepLabCut model variants are provided, trained on handheld field video of habituated wild gorillas collected in Loango National Park, Gabon (3840×2160, 50fps). Models vary by training dataset size and background segmentation method, allowing users to assess the effect of training data volume and subject isolation on pose estimation accuracy. All models predict 16 anatomical landmarks corresponding to those defined by OpenApePose: nose, eyes, head crown, neck, shoulders, elbows, wrists, hip, knees, and ankles, using a ResNet-50 backbone. Background-segmented variants were generated using the Segment Anything Model 2 (SAM2; Meta AI).

Repository contents
| DLC-S | DLC-M | DLC-L | DLC-GBG | DLC-ALPHA | DLC-OAP-BASE |
These folders contain the complete DeepLabCut project folders for each model variant, including labeled training data, model configuration files, and pose estimation predictions on novel videos. Note: DLC-OAP-BASE does not include training data; for the original training data see Desai et al. (2023).
| OAP |
Pose estimation predictions generated by the OpenApePose model for all novel videos, for comparative evaluation against the DeepLabCut variants.
| Novel_video_&_ground_truth |
24 novel videos of wild gorillas alongside manually labelled ground truth frames, used to evaluate and compare all model variants.
| Walking_stride_analysis |
DLC-L pose estimation predictions for 6 complete terrestrial knuckle-walking strides alongside ground truth data, used to assess landmark tracking accuracy during locomotion.

Licensing and attribution
The data, model weights, and outputs in this repository are released under CC BY 4.0. You are free to share and adapt this material provided appropriate credit is given and the associated publication is cited.
DeepLabCut is a third-party tool used under the GNU Lesser General Public License v3.0 (LGPL-3.0). Users of these models should ensure compliance with the DeepLabCut licence terms, available at: github.com
OpenApePose is a third-party tool used under its respective licence, available at: github.com. Users should consult this licence before redistributing OpenApePose outputs.

Please cite the following when using this repository:

King et al. (in preparation)- associated publication
Mathis et al. (2018) - DeepLabCut
Desai et al. (2024) - OpenApePose