Detect, track and classify objects with a custom classification model on Android

You can use ML Kit to detect and track objects in successive video frames.

When you pass an image to ML Kit, it detects up to five objects in the image along with the position of each object in the image. When detecting objects in video streams, each object has a unique ID that you can use to track the object from frame to frame.

To classify the objects that are detected, you must use a custom image classification model and bundle it with your app. Please refer to Custom models with ML Kit for guidance on where to find pre-trained models or how to train your own, including model compatibility requirements.

See the ML Kit Material Design showcase app and the ML Kit quickstart sample on GitHub for examples of this API in use.

Before you begin

  1. In your project-level build.gradle file, make sure to include Google's Maven repository in both your buildscript and allprojects sections.
  2. Add the dependencies for the ML Kit Android libraries to your module's app-level gradle file, which is usually app/build.gradle:
    dependencies {
      // ...
      // Object detection and tracking feature with custom classifier
      implementation 'com.google.mlkit:object-detection-custom:16.0.0'
    }
    

1. Bundle model with your app

To bundle your TensorFlow Lite model with your app, copy the model file (usually ending in .tflite or .lite) to your app's assets/ folder. (You might need to create the folder first by right-clicking the app/ folder, then clicking New > Folder > Assets Folder.)

Then, add the following to your app's build.gradle file to ensure Gradle doesn’t compress the models when building the app:

android {
    // ...
    aaptOptions {
        noCompress "tflite"  // Your model's file extension: "tflite", "lite", etc.
    }
}

The model file will be included in the app package and available to ML Kit as a raw asset.

2. Prepare the input image

Create an InputImage object from your image. The object detector runs directly from a Bitmap, NV21 ByteBuffer or a YUV_420_888 media.Image. Constructing an InputImage from those sources is recommended if you have direct access to one of them. If you construct an InputImage from other sources, we will handle the conversion internally for you and it might be less efficient.

You can create an InputImage from different sources, each is explained below.

Using a media.Image

To create an InputImage object from a media.Image object, such as when you capture an image from a device's camera, pass the media.Image object and the image's rotation to InputImage.fromMediaImage().

If you use the CameraX library, the OnImageCapturedListener and ImageAnalysis.Analyzer classes calculate the rotation value for you.

Kotlin

private class YourImageAnalyzer : ImageAnalysis.Analyzer {

    override fun analyze(imageProxy: ImageProxy) {
        val mediaImage = imageProxy.image
        if (mediaImage != null) {
            val image = InputImage.fromMediaImage(mediaImage, imageProxy.imageInfo.rotationDegrees)
            // Pass image to an ML Kit Vision API
            // ...
        }
    }
}

Java

private class YourAnalyzer implements ImageAnalysis.Analyzer {

    @Override
    public void analyze(ImageProxy imageProxy) {
        Image mediaImage = imageProxy.getImage();
        if (mediaImage != null) {
          InputImage image =
                InputImage.fromMediaImage(mediaImage, imageProxy.getImageInfo().getRotationDegrees());
          // Pass image to an ML Kit Vision API
          // ...
        }
    }
}

If you don't use a camera library that gives you the image's rotation degree, you can calculate it from the device's rotation degree and the orientation of camera sensor in the device:

Kotlin

private val ORIENTATIONS = SparseIntArray()

init {
    ORIENTATIONS.append(Surface.ROTATION_0, 0)
    ORIENTATIONS.append(Surface.ROTATION_90, 90)
    ORIENTATIONS.append(Surface.ROTATION_180, 180)
    ORIENTATIONS.append(Surface.ROTATION_270, 270)
}

/**
 * Get the angle by which an image must be rotated given the device's current
 * orientation.
 */
@RequiresApi(api = Build.VERSION_CODES.LOLLIPOP)
@Throws(CameraAccessException::class)
private fun getRotationCompensation(cameraId: String, activity: Activity, isFrontFacing: Boolean): Int {
    // Get the device's current rotation relative to its "native" orientation.
    // Then, from the ORIENTATIONS table, look up the angle the image must be
    // rotated to compensate for the device's rotation.
    val deviceRotation = activity.windowManager.defaultDisplay.rotation
    var rotationCompensation = ORIENTATIONS.get(deviceRotation)

    // Get the device's sensor orientation.
    val cameraManager = activity.getSystemService(CAMERA_SERVICE) as CameraManager
    val sensorOrientation = cameraManager
            .getCameraCharacteristics(cameraId)
            .get(CameraCharacteristics.SENSOR_ORIENTATION)!!

    if (isFrontFacing) {
        rotationCompensation = (sensorOrientation + rotationCompensation) % 360
    } else { // back-facing
        rotationCompensation = (sensorOrientation - rotationCompensation + 360) % 360
    }
    return rotationCompensation
}

Java

private static final SparseIntArray ORIENTATIONS = new SparseIntArray();
static {
    ORIENTATIONS.append(Surface.ROTATION_0, 0);
    ORIENTATIONS.append(Surface.ROTATION_90, 90);
    ORIENTATIONS.append(Surface.ROTATION_180, 180);
    ORIENTATIONS.append(Surface.ROTATION_270, 270);
}

/**
 * Get the angle by which an image must be rotated given the device's current
 * orientation.
 */
@RequiresApi(api = Build.VERSION_CODES.LOLLIPOP)
private int getRotationCompensation(String cameraId, Activity activity, boolean isFrontFacing)
        throws CameraAccessException {
    // Get the device's current rotation relative to its "native" orientation.
    // Then, from the ORIENTATIONS table, look up the angle the image must be
    // rotated to compensate for the device's rotation.
    int deviceRotation = activity.getWindowManager().getDefaultDisplay().getRotation();
    int rotationCompensation = ORIENTATIONS.get(deviceRotation);

    // Get the device's sensor orientation.
    CameraManager cameraManager = (CameraManager) activity.getSystemService(CAMERA_SERVICE);
    int sensorOrientation = cameraManager
            .getCameraCharacteristics(cameraId)
            .get(CameraCharacteristics.SENSOR_ORIENTATION);

    if (isFrontFacing) {
        rotationCompensation = (sensorOrientation + rotationCompensation) % 360;
    } else { // back-facing
        rotationCompensation = (sensorOrientation - rotationCompensation + 360) % 360;
    }
    return rotationCompensation;
}

Then, pass the media.Image object and the rotation degree value to InputImage.fromMediaImage():

Kotlin

val image = InputImage.fromMediaImage(mediaImage, rotation)

Java

InputImage image = InputImage.fromMediaImage(mediaImage, rotation);

Using a file URI

To create an InputImage object from a file URI, pass the app context and file URI to InputImage.fromFilePath(). This is useful when you use an ACTION_GET_CONTENT intent to prompt the user to select an image from their gallery app.

Kotlin

val image: InputImage
try {
    image = InputImage.fromFilePath(context, uri)
} catch (e: IOException) {
    e.printStackTrace()
}

Java

InputImage image;
try {
    image = InputImage.fromFilePath(context, uri);
} catch (IOException e) {
    e.printStackTrace();
}

Using a ByteBuffer or ByteArray

To create an InputImage object from a ByteBuffer or a ByteArray, first calculate the image rotation degree as previously described for media.Image input. Then, create the InputImage object with the buffer or array, together with image's height, width, color encoding format, and rotation degree:

Kotlin

val image = InputImage.fromByteBuffer(
        byteBuffer,
        /* image width */ 480,
        /* image height */ 360,
        rotationDegrees,
        InputImage.IMAGE_FORMAT_NV21 // or IMAGE_FORMAT_YV12
)
// Or:
val image = InputImage.fromByteArray(
        byteArray,
        /* image width */ 480,
        /* image height */ 360,
        rotationDegrees,
        InputImage.IMAGE_FORMAT_NV21 // or IMAGE_FORMAT_YV12
)

Java

InputImage image = InputImage.fromByteBuffer(byteBuffer,
        /* image width */ 480,
        /* image height */ 360,
        rotationDegrees,
        InputImage.IMAGE_FORMAT_NV21 // or IMAGE_FORMAT_YV12
);
// Or:
InputImage image = InputImage.fromByteArray(
        byteArray,
        /* image width */480,
        /* image height */360,
        rotation,
        InputImage.IMAGE_FORMAT_NV21 // or IMAGE_FORMAT_YV12
);

Using a Bitmap

To create an InputImage object from a Bitmap object, make the following declaration:

Kotlin

val image = InputImage.fromBitmap(bitmap, 0)

Java

InputImage image = InputImage.fromBitmap(bitmap, rotationDegree);

The image is represented by a Bitmap object together with rotation degrees.

3. Configure and run the object detector

  1. Create a new local model

Kotlin

val localModel =
        LocalModel.Builder()
        .setAssetFilePath(assetFilePathToTfliteModel)
        // or .setAbsoluteFilePath(absoluteFilePathToTfliteModel)
        .build()

Java

LocalModel localModel =
    new LocalModel.Builder()
        .setAssetFilePath(assetFilePathToTfliteModel)
        // or .setAbsoluteFilePath(absoluteFilePathToTfliteModel)
        .build();
  1. Configure the object detector for your use case with a CustomObjectDetectorOptions object. You can change the following settings:

    Object Detector Settings
    Detection mode STREAM_MODE (default) | SINGLE_IMAGE_MODE

    In STREAM_MODE (default), the object detector runs with low latency, but might produce incomplete results (such as unspecified bounding boxes or category labels) on the first few invocations of the detector. Also, in STREAM_MODE, the detector assigns tracking IDs to objects, which you can use to track objects across frames. Use this mode when you want to track objects, or when low latency is important, such as when processing video streams in real time.

    In SINGLE_IMAGE_MODE, the object detector returns the result after the object's bounding box is determined. If you also enable classification it returns the result after the bounding box and category label are both available. As a consequence, detection latency is potentially higher. Also, in SINGLE_IMAGE_MODE, tracking IDs are not assigned. Use this mode if latency isn't critical and you don't want to deal with partial results.

    Detect and track multiple objects false (default) | true

    Whether to detect and track up to five objects or only the most prominent object (default).

    Classify objects false (default) | true

    Whether or not to classify detected objects by using the provided custom classifier model. To use your custom classification model, you need to set this to true.

    Classification confidence threshold

    Minimum confidence score of detected labels. If not set, any classifier threshold specified by the model’s metadata will be used. If the model does not contain any metadata or the metadata does not specify a classifier threshold, a default threshold of 0.0 will be used.

    Maximum labels per object

    Maximum number of labels per object that the detector will return. If not set, the default value of 10 will be used.

    The object detection and tracking API is optimized for these two core use cases:

    • Live detection and tracking of the most prominent object in the camera viewfinder.
    • The detection of multiple objects from a static image.

    To configure the API for these use cases:

Kotlin

// Live detection and tracking
val customObjectDetectorOptions =
        CustomObjectDetectorOptions.Builder(localModel)
        .setDetectorMode(CustomObjectDetectorOptions.STREAM_MODE)
        .enableClassification()
        .setClassificationConfidenceThreshold(0.5f)
        .setMaxPerObjectLabelCount(3)
        .build()

// Multiple object detection in static images
val customObjectDetectorOptions =
        CustomObjectDetectorOptions.Builder(localModel)
        .setDetectorMode(CustomObjectDetectorOptions.SINGLE_IMAGE_MODE)
        .enableMultipleObjects()
        .enableClassification()
        .setClassificationConfidenceThreshold(0.5f)
        .setMaxPerObjectLabelCount(3)
        .build()

Java

// Live detection and tracking
CustomObjectDetectorOptions customObjectDetectorOptions =
        new CustomObjectDetectorOptions.Builder(localModel)
                .setDetectorMode(CustomObjectDetectorOptions.STREAM_MODE)
                .enableClassification()
                .setClassificationConfidenceThreshold(0.5f)
                .setMaxPerObjectLabelCount(3)
                .build();

// Multiple object detection in static images
CustomObjectDetectorOptions customObjectDetectorOptions =
        new CustomObjectDetectorOptions.Builder(localModel)
                .setDetectorMode(CustomObjectDetectorOptions.SINGLE_IMAGE_MODE)
                .enableMultipleObjects()
                .enableClassification()
                .setClassificationConfidenceThreshold(0.5f)
                .setMaxPerObjectLabelCount(3)
                .build();
  1. Create a new object detector:

Kotlin

val objectDetector =
    ObjectDetection.getClient(customObjectDetectorOptions)

Java

ObjectDetector objectDetector =
    ObjectDetection.getClient(customObjectDetectorOptions);
  1. Use the detector:

Kotlin

objectDetector
    .process(image)
    .addOnFailureListener(e -> {...})
    .addOnSuccessListener(results -> {
        for (detectedObject in results) {
          // ...
        }
    });

Java

objectDetector
    .process(image)
    .addOnFailureListener(e -> {...})
    .addOnSuccessListener(results -> {
        for (DetectedObject detectedObject : results) {
          // ...
        }
    });

4. Get information about labeled objects

If the call to process() succeeds, a list of DetectedObjects is passed to the success listener.

Each DetectedObject contains the following properties:

Bounding box A Rect that indicates the position of the object in the image.
Tracking ID An integer that identifies the object across images. Null in SINGLE_IMAGE_MODE.
Labels
Label description The label's text description. Only returned if the TensorFlow Lite model's metadata contains label descriptions.
Label index The label's index among all the labels supported by the classifier.
Label confidence The confidence value of the object classification.

Kotlin

// The list of detected objects contains one item if multiple
// object detection wasn't enabled.
for (detectedObject in results) {
    val boundingBox = detectedObject.boundingBox
    val trackingId = detectedObject.trackingId
    for (label in detectedObject.labels) {
      val text = label.text
      val index = label.index
      val confidence = label.confidence
    }
}

Java

// The list of detected objects contains one item if multiple
// object detection wasn't enabled.
for (DetectedObject detectedObject : results) {
  Rect boundingBox = detectedObject.getBoundingBox();
  Integer trackingId = detectedObject.getTrackingId();
  for (Label label : detectedObject.getLabels()) {
    String text = label.getText();
    int index = label.getIndex();
    float confidence = label.getConfidence();
  }
}

Ensuring a great user experience

For the best user experience, follow these guidelines in your app:

  • Successful object detection depends on the object's visual complexity. In order to be detected, objects with a small number of visual features might need to take up a larger part of the image. You should provide users with guidance on capturing input that works well with the kind of objects you want to detect.
  • When you use classification, if you want to detect objects that don't fall cleanly into the supported categories, implement special handling for unknown objects.

Also, check out the ML Kit Material Design showcase app and the Material Design Patterns for machine learning-powered features collection.

Improving performance

If you want to use object detection in a real-time application, follow these guidelines to achieve the best framerates:

  • When you use streaming mode in a real-time application, don't use multiple object detection, as most devices won't be able to produce adequate framerates.

  • If you use the Camera or camera2 API, throttle calls to the detector. If a new video frame becomes available while the detector is running, drop the frame. See the VisionProcessorBase class in the quickstart sample app for an example.
  • If you use the CameraX API, be sure that backpressure strategy is set to its default value ImageAnalysis.STRATEGY_KEEP_ONLY_LATEST. This guarantees only one image will be delivered for analysis at a time. If more images are produced when the analyzer is busy, they will be dropped automatically and not queued for delivery. Once the image being analyzed is closed by calling ImageProxy.close(), the next latest image will be delivered.
  • If you use the output of the detector to overlay graphics on the input image, first get the result from ML Kit, then render the image and overlay in a single step. This renders to the display surface only once for each input frame. See the CameraSourcePreview and GraphicOverlay classes in the quickstart sample app for an example.
  • If you use the Camera2 API, capture images in ImageFormat.YUV_420_888 format. If you use the older Camera API, capture images in ImageFormat.NV21 format.

Next steps

See the ML Kit Material Design showcase app and the ML Kit quickstart sample on GitHub for examples of this API in use.