OrangIdentifier

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

Source code (ML Pipeline): github.com/tit-exe/OrangIdentifier
Source code (Android App): github.com/tit-exe/OrangIdentifier-Android

Overview

This pipeline trains a face detector and an individual identification model from labeled photographs, then exports the result as a lightweight gallery JSON for an Android app that runs entirely offline.

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.

Android App Assets

This repository also hosts the exported .tflite models and the gallery database required to run the offline Android app.

If you are building the Android app, download these three files from the main branch and place them in the app/src/main/assets/ folder of the Android project:

• gallery.json: The database containing the identity prototypes.
• yolo_v2_detector.tflite: The optimized YOLO face detector model.
• megadesc_T_arcface_backbone.tflite: The MegaDescriptor-T embedding backbone (112 MB).

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)

Models

File	Used in	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_40individuals_acc99.pt	V4 ★	105 MB	ArcFace, 40 individuals

Performance

	V1	V2	V3	V4
Backbone	ResNet50	ResNet50	MegaDescriptor-T	MegaDescriptor-T
Supervised individuals	10	10	10	40
Closed-set accuracy	96.3%	~98%	99.2%	99.2%
Unknown rejection (1,622 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
Captive collection	10	2,127	Training (known)
Field rescue center	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_40individuals_acc99.pt"
)

Or via the pipeline:

python models/download_models.py

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.
• Deng et al. (2019). ArcFace. CVPR 2019.
• Deng et al. (2020). Sub-center ArcFace. ECCV 2020.
• Liu et al. (2021). Swin Transformer. ICCV 2021.
• Jocher et al. (2023). Ultralytics YOLOv8.