OrangIdentifier

Individual facial recognition for Bornean orangutans — end-to-end pipeline from raw photographs to offline Android deployment.

CNRS IPHC Strasbourg · BOS Foundation Borneo · May 2026

---

Overview

Rangers at BOS Foundation (Borneo Orangutan Survival) need to identify individual orangutans during field patrols. This pipeline produces an Android app that runs entirely offline and returns an identity match or "unknown individual" within seconds of taking a photo.

The gallery is a JSON file containing one averaged embedding vector per individual. Adding a new individual requires 10–20 photos, takes under a minute, and requires no retraining.

---

Models

File	Version	Size	Description
yolo_v1_nano_mAP92.pt	V1	6 MB	YOLO nano — mAP@50 = 91.98%
yolo_v2_medium_mAP99.pt	V1–V4	85 MB	YOLO medium — mAP@50 = 99.39%
resnet50_classifier_10classes_acc96.pt	V1	90 MB	Closed-set classifier, acc = 96.3%
resnet50_backbone_2048dim.pt	V2	90 MB	Embedding backbone, 2048-dim
megadesc_T_arcface_final_epoch21_acc99.pt	V3	105 MB	ArcFace, 10 individuals
megadesc_T_arcface_v4_40individus_acc99.pt	V4 ★	105 MB	ArcFace, 40 individuals

---

Inference pipeline

Raw photo → YOLO face detection (mAP@50=99.4%) → 224×224 crop
→ MegaDescriptor-T-224 (Swin Transformer, 768-dim embedding)
→ Cosine similarity vs gallery
→ Known individual (sim ≥ 0.22) or Unknown (sim < 0.22)

---

Performance

	V1	V2	V3	V4
Backbone	ResNet50	ResNet50	MegaDescriptor-T	MegaDescriptor-T
Supervised individuals	10	10	10	40
Zoo accuracy	96.3%	~98%	99.2%	99.2%
BOS rejection (1622 unseen crops)	—	27.5%	97.5%	97.5%
Wild internet rejection	—	48.5%	93.2%	93.0%
Separability gap	—	0.294	0.883	0.885

---

Dataset

Source	Individuals	Crops	Role
Zoo Amnéville + Indonesia	10	2,127	Training (known)
BOS Foundation Borneo	30	1,622	Open-set test only
Internet (iNaturalist, GBIF, web)	unlabeled	5,429	Background class

Images are not included.

---

Download

from huggingface_hub import hf_hub_download

path = hf_hub_download(
    repo_id="tit0000/OrangIdentifier",
    filename="megadesc_T_arcface_v4_40individus_acc99.pt"
)

---

Source code & documentation

github.com/tit0000/OrangIdentifier

---

Security note

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

---

References

• Čermák et al. (2024). WildlifeDatasets. WACV 2024. CVF
• Deng et al. (2019). ArcFace. CVPR 2019. arXiv
• Deng et al. (2020). Sub-center ArcFace. ECCV 2020. Springer
• Liu et al. (2021). Swin Transformer. ICCV 2021. arXiv
• Otarashvili, L. (2023). MiewID. Conservation X Labs. GitHub
• Jocher et al. (2023). Ultralytics YOLOv8. GitHub