Skip to content

v4

Latest
Latest

Choose a tag to compare

View all tags
@tit-exe tit-exe released this 26 Jun 13:36
· 4 commits to main since this release
v4
835aedc

OrangIdentifier Android - V4.0

Release of the 4th major version of the OrangIdentifier Android companion app. This offline mobile application allows field users to photograph Bornean orangutans and identify them in real-time, completely without internet access.

Historical context: V1 used a ResNet closed-set classifier, V2 moved to ResNet embeddings, V3 introduced MegaDescriptor with a single centroid per individual. This V4 release integrates the V6 AI model and upgrades the gallery format and scoring system accordingly, using SubCenterArcFace training and a multi-exemplar representation for significantly improved recognition accuracy.

🌟 Key Changes from V3

  • Multi-exemplar gallery (model V6): Each individual is now represented by 25 exemplar vectors (768-dim each) instead of a single centroid. The recognition score is computed as max(dot(query, exemplar_i)) over all 25 exemplars, capturing natural appearance variability across angles and lighting conditions far better than an average vector.
  • New gallery: 15 known individuals, calibrated threshold of 0.5371, separability gap of 0.6711.
  • Gallery management: Full gallery export, per-individual patch sharing, automatic versioning with up to 20 restore points, and one-tap undo.
  • Field photos: Add photos to an existing individual's profile in the field. Stored separately from the training exemplars and merged at inference time.
  • Model hot-swap: Import a new backbone or detector as a .zip bundle via Settings without reinstalling the app.

🌟 Features

  • 100% Offline Inference: Runs entirely on-device using optimized TensorFlow Lite models.
  • V6 AI Pipeline (MegaDescriptor-T-224 + SubCenterArcFace): YOLO v2 face detector + Swin Transformer backbone producing 768-dim embeddings, matched against 25 exemplars per individual for open-set recognition.
  • Instant Onboarding: Add new individuals directly from the app with at least 5 photos. The prototype is computed and saved to the local gallery immediately, no retraining required.
  • Gallery Sharing: Export the full gallery or individual patches as JSON files and share them with other devices.
  • Automatic Backups: Every gallery modification creates a backup. Full restore history accessible from Settings.
  • Scan History: A local Room database keeps a log of all past identifications.

📥 Installation

  1. Download the app-release.apk attached to this release under Assets.
  2. Open the APK file on your Android device to install it. You may need to enable Install from unknown sources in your Android settings.

⚠️ Note for Developers

All models and the gallery are included directly in the repository. Clone the project, open it in Android Studio, sync Gradle, and build.

Assets 3
Loading