This project trains a face detector and an individual identification model from labeled field photographs, then exports the result as a lightweight gallery JSON for an Android app that runs entirely offline. The gallery holds up to 25 exemplar embeddings per individual. Adding a new individual takes a handful of photos on the phone and requires no retraining.

Inference pipeline

Field photo -> YOLO gorilla face detection
            -> 224x224 crop
            -> MegaDescriptor-T-224 (Swin Transformer Tiny, 768-dim embedding)
            -> max cosine similarity over the exemplars of each individual
            -> Known individual (score >= 0.4689 and margin >= 0.08) or Unknown

Android app assets

This repository hosts the assets required to run the offline Android app. The app identifies files by role, so the recognition backbone must be downloaded here (it exceeds the GitHub 100 MB limit), while the detector and the gallery are also bundled in the app repository:

megadesc_T_arcface_backbone.tflite : the MegaDescriptor-T embedding backbone (107 MB). Download it and place it in app/src/main/assets/ before building the app.
yolo_v2_detector.tflite : the gorilla face detector (the filename is the one the Android app expects; it is the gorilla detector, not an orangutan model).
gallery.json : the identity database, 66 individuals.

Models

File	Role	Size	Description
`yolo_gorilla.pt`	pipeline	18 MB	Gorilla face detector (YOLOv8), used for crop extraction and training
`gorilla_v1_best.pt`	pipeline	105 MB	Trained V1 identifier checkpoint (MegaDescriptor-T + Sub-center ArcFace)
`megadesc_T_arcface_backbone.tflite`	app	107 MB	Identifier backbone exported to TFLite for the Android app
`yolo_v2_detector.tflite`	app	6 MB	Gorilla face detector exported to TFLite for the Android app
`gallery.json`	app	30 MB	Identity gallery, 66 individuals, up to 25 exemplars each, 768-dim

The generic MegaDescriptor-T-224 backbone used as the training starting point is not stored here. timm downloads it automatically from BVRA/MegaDescriptor-T-224 the first time training runs.

Performance

Version 1, 66 individuals, Virunga 2025. Metrics are measured on the held-out validation set after training.

Metric	Value
Recognized individuals	66
Top-1 accuracy	93.0%
Top-3 accuracy	96.1%
Mean F1	0.981
Composite score	0.808
Rejection threshold	0.4689
Separability gap	0.4351
Backbone	MegaDescriptor-T-224 (Swin Transformer Tiny, 27.5M parameters)
Training time	about 66 minutes on an RTX 3050 4 GB

The rejection threshold is the cosine-similarity cutoff below which a face is reported as unknown, calibrated by maximizing F1 on the validation set. The separability gap is the average similarity gap between an individual's own exemplars and its closest rival; a higher gap means less confusion.

Dataset

Source	Individuals	Crops	Role
Field photographs (Virunga)	66 known (+ 3 held out)	2,809	Training and validation
Internet / background images	unlabeled	428	Background class (pseudo-unknowns)

Two individuals with too few crops were excluded from training, and three were held out as pseudo-unknowns to calibrate the rejection threshold. Photographs are not included in this repository.

Download

python

from huggingface_hub import hf_hub_download

path = hf_hub_download(
    repo_id="tit0000/GorillaIdentifier",
    filename="gorilla_v1_best.pt",
)

Or, for the pipeline detector, via the helper script in the code repository:

python models/download_models.py

Security note

These .pt files are standard PyTorch and Ultralytics checkpoints. The pickle imports flagged by Hugging Face come from trusted libraries (torch, ultralytics, collections) and contain no malicious code.

References

Čermák et al. (2024). WildlifeDatasets. WACV 2024.
Deng et al. (2019). ArcFace. CVPR 2019.
Deng et al. (2020). Sub-center ArcFace. ECCV 2020.
Liu et al. (2021). Swin Transformer. ICCV 2021.
Khosla et al. (2020). Supervised Contrastive Learning. NeurIPS 2020.