An Android app that runs on-device object detection using LiteRT (TensorFlow Lite Runtime). Pick any photo from your gallery — the app detects objects, draws labelled bounding boxes, and shows the result alongside the original image. No internet connection required.
| Home | Photo picker | Result — crowd | Result — dog & bike |
|---|---|---|---|
 |
 |
 |
 |
The project follows Clean Architecture with a strict unidirectional dependency rule:
Presentation → Domain ← Data
Each layer only depends inward. The domain layer has zero Android dependencies.
┌─────────────────────────────────────────────────────────┐
│ Presentation (Compose + ViewModel) │
│ HomeScreen → picks image │
│ ResultScreen → shows original + annotated image │
└────────────────────┬────────────────────────────────────┘
│ calls use cases
┌────────────────────▼────────────────────────────────────┐
│ Domain (pure Kotlin) │
│ LoadImageUseCase │
│ DetectObjectsUseCase │
│ DrawBoundingBoxesUseCase │
│ DetectionRepository / ImageRepository / BitmapRenderer│
└────────────────────┬────────────────────────────────────┘
│ implemented by
┌────────────────────▼────────────────────────────────────┐
│ Data │
│ TfliteObjectDetector ← LiteRT interpreter │
│ DetectionRepositoryImpl │
│ ImageRepositoryImpl ← ContentResolver │
│ BitmapRendererImpl ← Canvas drawing │
└─────────────────────────────────────────────────────────┘
android_object_detection/
├── gradle/
│ └── libs.versions.toml # centralised version catalog
├── app/
│ ├── build.gradle.kts
│ └── src/
│ ├── main/
│ │ ├── assets/
│ │ │ ├── efficientdet_lite0.tflite # SSD MobileNet V1 COCO quantized
│ │ │ └── coco_labels.txt # 80 COCO class labels
│ │ ├── AndroidManifest.xml
│ │ └── java/com/example/android_object_detection/
│ │ │
│ │ ├── core/
│ │ │ └── util/
│ │ │ ├── Result.kt # sealed interface: Success / Error / Loading
│ │ │ └── ErrorMessages.kt # string constants for error states
│ │ │
│ │ ├── domain/
│ │ │ ├── model/
│ │ │ │ ├── DetectionResult.kt # label + confidence + bounding box
│ │ │ │ └── BoundingBox.kt # left / top / right / bottom (pixels)
│ │ │ ├── repository/
│ │ │ │ ├── DetectionRepository.kt
│ │ │ │ ├── ImageRepository.kt
│ │ │ │ └── BitmapRenderer.kt
│ │ │ └── usecase/
│ │ │ ├── LoadImageUseCase.kt
│ │ │ ├── DetectObjectsUseCase.kt
│ │ │ └── DrawBoundingBoxesUseCase.kt
│ │ │
│ │ ├── data/
│ │ │ ├── detector/
│ │ │ │ ├── ObjectDetector.kt # interface: detect(Bitmap)
│ │ │ │ └── TfliteObjectDetector.kt # LiteRT implementation
│ │ │ └── repository/
│ │ │ ├── DetectionRepositoryImpl.kt
│ │ │ ├── ImageRepositoryImpl.kt
│ │ │ └── BitmapRendererImpl.kt
│ │ │
│ │ ├── di/
│ │ │ └── AppModule.kt # Hilt bindings
│ │ │
│ │ ├── presentation/
│ │ │ ├── components/
│ │ │ │ └── PrimaryButton.kt
│ │ │ ├── home/
│ │ │ │ ├── HomeScreen.kt
│ │ │ │ ├── HomeViewModel.kt
│ │ │ │ ├── HomeUiState.kt
│ │ │ │ └── HomeNavEntry.kt
│ │ │ ├── result/
│ │ │ │ ├── ResultScreen.kt
│ │ │ │ ├── ResultViewModel.kt
│ │ │ │ ├── ResultUiState.kt
│ │ │ │ └── ResultNavEntry.kt
│ │ │ ├── navigation/
│ │ │ │ ├── AppNavGraph.kt
│ │ │ │ ├── Navigator.kt
│ │ │ │ └── Route.kt
│ │ │ └── theme/
│ │ │ └── AppTheme.kt
│ │ │
│ │ ├── ObjectDetectionApp.kt # @HiltAndroidApp
│ │ └── MainActivity.kt # @AndroidEntryPoint
│ │
│ └── test/
│ └── java/com/example/android_object_detection/
│ ├── data/
│ │ ├── detector/
│ │ │ └── FakeObjectDetector.kt # test double
│ │ └── repository/
│ │ └── DetectionRepositoryImplTest.kt
│ └── domain/
│ └── usecase/
│ └── DetectObjectsUseCaseTest.kt
| Concern | Library |
|---|---|
| Language | Kotlin 2.3.20 |
| UI | Jetpack Compose + Material3 |
| Navigation | Navigation3 1.1.0 |
| Dependency injection | Hilt 2.59.2 |
| Object detection | LiteRT (TFLite) 2.1.4 |
| Image loading | Coil 2.7.0 |
| Async | Kotlin Coroutines + Flow |
| Testing | JUnit4 · Mockk · Turbine · kotlinx-coroutines-test |
| Build | AGP 9.1.1 · Gradle version catalog |
URI string
│
▼ LoadImageUseCase
ByteArray (raw image bytes)
│
▼ DetectObjectsUseCase
│ └── DetectionRepositoryImpl
│ └── TfliteObjectDetector
│ ├── resize bitmap → 300×300
│ ├── extract RGB pixels → ByteBuffer (UINT8)
│ └── Interpreter.runForMultipleInputsOutputs()
│ outputs: boxes · classes · scores · count
│
List<DetectionResult>
│
▼ DrawBoundingBoxesUseCase
│ └── BitmapRendererImpl (Canvas)
│
ByteArray (annotated PNG)
│
▼ ResultScreen
| Property | Value |
|---|---|
| File | assets/efficientdet_lite0.tflite |
| Architecture | SSD MobileNet V1 |
| Dataset | COCO (80 classes) |
| Input | 300 × 300 RGB UINT8 |
| Max detections | 10 |
| Confidence threshold | 40% |
| Source | TFLite object detection example |
- Android Studio Meerkat or newer
- Android device / emulator running API 26+
- JDK 17
git clone <repo-url>
cd android_object_detection
./gradlew installDebug./gradlew testDebugUnitTestWhy LiteRT instead of ML Kit? LiteRT gives direct control over model selection, input preprocessing, and output parsing. ML Kit is a convenience wrapper with limited flexibility and requires Google Play Services on the device.
Why a separate ObjectDetector interface?
DetectionRepositoryImpl depends on the interface, not the concrete TFLite class. This makes the
repository fully unit-testable with FakeObjectDetector — no Android runtime needed.
Why Dispatchers.Default for inference?
TFLite inference is CPU-bound, not I/O-bound. Dispatchers.Default uses a thread pool sized to the
number of CPU cores, which is the right scheduler for compute work.
Why is TfliteObjectDetector a singleton?
The Interpreter is expensive to construct (model file mapping + JNI initialisation). Making it a
singleton means it is created once on first use and reused for every detection call.