dog_detection library

On-device dog detection and landmark estimation using TensorFlow Lite.

This library provides a Flutter plugin for dog detection using a unified multi-stage TFLite pipeline: SSD body detection, species classification, body pose estimation, face localization, and face landmark extraction.

Quick Start:

import 'package:dog_detection/dog_detection.dart';

final detector = DogDetector(mode: DogDetectionMode.full);
await detector.initialize();

final dogs = await detector.detect(imageBytes);
for (final dog in dogs) {
  print('Dog at ${dog.boundingBox} score=${dog.score}');
  if (dog.pose != null) {
    final tail = dog.pose!.getLandmark(AnimalPoseLandmarkType.tailEnd);
    print('Tail: (${tail?.x}, ${tail?.y})');
  }
  if (dog.face != null && dog.face!.hasLandmarks) {
    final nose = dog.face!.getLandmark(DogLandmarkType.noseBridgeBottom);
    print('Nose: (${nose?.x}, ${nose?.y})');
  }
}

await detector.dispose();

Main Classes:

  • DogDetector: Main API for dog detection. Runs the whole pipeline in a background isolate it owns, so detection never blocks the UI thread.
  • Dog: Top-level detection result with body, pose and face info
  • DogFace: Detected dog face with bounding box and landmarks
  • DogLandmark: Single face keypoint with 2D coordinates
  • BoundingBox: Axis-aligned rectangle in pixel coordinates

Detection Modes:

Pose Model Variants:

Face Landmark Model Variants:

  • DogLandmarkModel.full: Single model at 384px input resolution (bundled)
    • flip TTA (18 passes). Extra models downloaded on-demand from GitHub Releases (~110MB)

Skeleton Connections:

Classes

Animal
Top-level result for a single detected animal.
AnimalDetector
On-device animal detection using a multi-stage LiteRT pipeline.
AnimalPose
Full-body pose result for a single detected animal.
AnimalPoseLandmark
A single body pose keypoint with 2D coordinates and a confidence score.
BoundingBox
An axis-aligned or rotated bounding box defined by four corner points.
CameraFrame
A camera frame packaged for off-thread colour conversion and inference.
CropMetadata
Metadata for mapping landmark coordinates from crop space back to original image space.
Dog
Top-level result for a single detected dog.
DogDetectionDart
Dart plugin registration for dog_detection.
DogDetector
On-device dog detection using a unified multi-stage TensorFlow Lite pipeline.
DogFace
Detected dog face with bounding box and optional landmarks.
DogLandmark
A single dog face keypoint with 2D coordinates.
FpsCounter
A simple 1-second rolling FPS counter for camera-preview apps.
FrameThrottle
A single-slot gate for camera-frame processing that drops frames arriving while a previous frame is still being processed.
Mat
ModelDownloader
Downloads and caches the HRNet body pose model from GitHub Releases.
OneEuroFilter
One Euro filter for low-latency smoothing of a noisy 1D signal sampled at an irregular rate.
PerformanceConfig
Configuration for interpreter hardware acceleration and threading.
Point
A point with x, y, and optional z coordinates.

Enums

Accelerator
Hardware accelerator requested for LiteRT Next compilation.
AnimalPoseLandmarkType
SuperAnimal body keypoint types (indices 15-38 in the full SuperAnimal topology).
AnimalPoseModel
Body pose model variant for SuperAnimal keypoint extraction.
CameraFrameConversion
The colour conversion a CameraFrame's bytes need before being used as a 3-channel BGR image. Detector packages map this to an opencv COLOR_* code at the point of decode, inside their existing detection isolate.
CameraFrameRotation
Optional rotation applied after colour conversion. Detector packages map this to an opencv ROTATE_* code.
DogDetectionMode
Detection mode controlling the full pipeline behavior.
DogLandmarkModel
Dog landmark model variant for landmark extraction.
DogLandmarkType
Dog face landmark types based on the DogFLW dataset topology.
PerformanceMode
Hardware acceleration mode for LiteRT inference.
Precision
GPU precision mode for LiteRT Next compilation.

Constants

animalPoseConnections → const List<List<AnimalPoseLandmarkType>>
Defines the standard skeleton connections between SuperAnimal body keypoints.
dogLandmarkConnections → const List<List<DogLandmarkType>>
Defines the standard skeleton connections between dog face landmarks.
dogLandmarkFlipIndex → const List<int>
Landmark index permutation for horizontal flip (DogFLW convention).
IMREAD_COLOR → const int
numDogLandmarks → const int
Number of dog face landmarks (46 for the DogFLW model).

Functions

barQuarterTurns(DeviceOrientation orientation) → int
Quarter-turns (clockwise) to rotate a top-bar widget so it reads upright when the device is in landscape. Use with RotatedBox(quarterTurns: ...).
detectionSize({required int width, required int height, required CameraFrameRotation? rotation, required int maxDim}) → Size
Compute the final detection-image size used by overlay painters to map detector coordinates back onto the widget coord space.
imdecode(Uint8List buf, int flags, {Mat? dst}) → Mat
imdecode reads an image from a buffer in memory. The function imdecode reads an image from the specified buffer in memory. If the buffer is too short or contains invalid data, the function returns an empty matrix. @param buf Input array or vector of bytes. @param flags The same flags as in cv::imread, see cv::ImreadModes.
iouLTRB(double aLeft, double aTop, double aRight, double aBottom, double bLeft, double bTop, double bRight, double bBottom) → double
Exact intersection-over-union of two axis-aligned boxes in left/top/right/ bottom coordinates. Returns 0.0 when the boxes are degenerate or disjoint.
prepareCameraFrame({required int width, required int height, required List<CameraPlane> planes, CameraFrameRotation? rotation, bool isBgra = true}) → CameraFrame?
Prepare a CameraFrame descriptor from raw camera planes, for use with a detector package's detectFromCameraFrame(...) method.
prepareCameraFrameFromImage(Object cameraImage, {CameraFrameRotation? rotation, bool? isBgra}) → CameraFrame?
Convenience wrapper around prepareCameraFrame that accepts any object duck-typed to package:camera's CameraImage (i.e. exposing width, height, and a planes iterable of objects with bytes, bytesPerRow, and bytesPerPixel getters).
rotationForFrame({required int width, required int height, required int sensorOrientation, required bool isFrontCamera, required DeviceOrientation deviceOrientation}) → CameraFrameRotation?
Compute the rotation needed to present a camera frame upright to an on-device detection model, given the camera's sensor orientation and the device's current physical orientation.