ultralytics_yolo 0.1.43

ultralytics_yolo: ^0.1.43 copied to clipboard

• Enhancement: Unify classification output format across all platforms to use official Results.summary() format
  • Android: Changed classification output to follow official Ultralytics Results.summary() format with name, class, confidence fields
  • iOS: Unified classification output to use name, class, confidence instead of topClass, topConfidence
  • Format: Classification results now follow the official format: {name: String, class: Int, confidence: Double, top5: List}
• Bug Fix: Fix Android CLASSIFY FloatArray compatibility issue by adding .toList() conversion for top5Confs to ensure Iterable compatibility when handling FloatArray and fix array bounds issues in Android classification top5 list iteration

• Feature: Add lensFacing parameter to YOLOView for default camera selection
  • Added LensFacing enum with back (default) and front values
  • Added lensFacing parameter to YOLOView widget (defaults to LensFacing.back)
  • Android: Added setLensFacing() method to YOLOView class
  • Android: Reads lensFacing from creation params and sets camera facing before initialization
  • iOS: Modified YOLOView init to accept cameraPosition parameter
  • iOS: Reads lensFacing from creation params and passes correct camera position to init
  • Usage: YOLOView(lensFacing: LensFacing.front) now starts with front camera by default
  • Backward Compatible: Defaults to back camera if not specified, maintaining existing behavior
  • Example: Updated example app controller to support lensFacing parameter

• Bug Fix: Fix MissingPluginException for default YOLO instances
  • Fixed critical issue where creating YOLO instances without useMultiInstance: true resulted in MissingPluginException: No implementation found for method predictSingleImage on channel yolo_single_image_channel_default
  • Modified ChannelConfig.createChannel() to treat 'default' instance ID as null for backward compatibility
  • The native side registers the default channel as yolo_single_image_channel without any suffix
  • Now properly handles default instances by not appending _default suffix to channel name
  • Upgrade Note: If you experienced MissingPluginException errors with v0.1.38+, this version resolves the issue

• Bug Fix: Fix double overlay rendering when using onResult callback with native overlays enabled
  • Root Cause: When showOverlays: true and onResult callback was set, both native and Flutter overlays were rendering simultaneously, causing duplicate bounding boxes
  • Flutter: Added safeguard to clear existing _currentDetections when double-overlay scenario is detected
  • Behavior: Preserves original showOverlays contract (controls Flutter overlay) while preventing double rendering when native overlays are also enabled
  • Impact: Eliminates duplicate bounding boxes, improves performance by reducing unnecessary setState calls, and maintains full callback functionality without breaking existing API contract

• Bug Fix: Fix showOverlays parameter not hiding native platform bounding boxes
  • Flutter: Added showOverlays parameter to creation parameters passed to native code
  • Flutter: Added setShowOverlays() method to YOLOViewController for dynamic overlay control
  • iOS: Added showOverlays property to YOLOView class with getter/setter
  • iOS: Modified bounding box rendering logic to respect showOverlays parameter
  • iOS: Added setShowOverlays method channel handler in SwiftYOLOPlatformView
  • Android: Added showOverlays property and setShowOverlays() method to YOLOView class
  • Android: Modified onDraw() method to check showOverlays before drawing overlays
  • Android: Added setShowOverlays method channel handler in YOLOPlatformView
  • Usage: YOLOView(showOverlays: false) now properly hides both Flutter and native overlays
  • Dynamic Control: Use yoloController.setShowOverlays(false) to toggle overlays at runtime

• Bug Fix: iOS performance metrics not updating in YOLOView

  • Moved EventChannel subscription from initState to _onPlatformViewCreated to ensure native channel readiness on iOS
  • Aligned streaming config key with iOS by renaming throttleInterval to throttleIntervalMs when sending params
  • iOS now sources performance metrics from the latest inference result: processingTimeMs = result.speed * 1000, fps = result.fps
  • Added fallback to YOLOStreamConfig.DEFAULT when stream config is nil on iOS
  • Minor cleanup in SwiftYOLOPlatformView.swift for streamlined event dispatch

• Example App

  • Camera Inference Screen fixes for Detection and FPS

• Critical Update: Add Android 16KB page size support for Android 15+ compatibility
  • Upgraded to NDK r28 for automatic 16KB page size alignment
  • Updated Google AI Edge LiteRT (TensorFlow Lite) from 1.2.0 to 1.4.0
  • Ensured compatibility with Google Play's November 2025 requirement
  • Native libraries now support both 4KB (current) and 16KB (future) page sizes
  • Updated Android Gradle Plugin to 8.5.2 and Gradle to 8.7

• New Feature: implement UI controls on Android platform

• Bug Fix: Fix iOS numItemsThreshold inconsistency with Android

• Bug Fix: Fix originalImage being null in onStreamingData callback
  • Fixed issue where originalImage was consistently null despite includeOriginalImage: true being set
  • Capture original image data in BasePredictor before clearing currentBuffer

• Bug Fix: Fix setState callback issue in YOLOView
  • Fixed issue where onResult and onStreamingData callbacks would stop working after setState calls
  • Improved didUpdateWidget logic to prevent unnecessary subscription recreation when callbacks are functionally equivalent
  • Added subscription existence check to avoid recreating working subscriptions
  • Fixed Android compilation errors related to YOLOStreamConfig and method access

• Bug Fix: Fix Android crash in YOLOInstanceManager
  • Added compatibility methods for better API consistency

• New Feature: Add useGpu parameter for GPU acceleration control
  • Added useGpu parameter to YOLO constructor and YOLOView widget
  • Allows disabling GPU acceleration on problematic devices to prevent crashes
  • Supports fallback to CPU/NNAPI on Android and CPU-only mode on iOS
  • Cross-platform consistency: parameter works on both Android and iOS
  • Default value is true for backward compatibility
  • Added comprehensive example demonstrating GPU control and error handling
  • Updated documentation with usage examples and troubleshooting guide

API Changes:

• YOLO constructor now accepts optional useGpu: bool = true parameter
• YOLOView widget now accepts optional useGpu: bool = true parameter
• YOLO.withClassifierOptions constructor now accepts optional useGpu: bool = true parameter

• Remove pubspec models

• Bug Fix: Clarified model placement requirements for Android
  • For Android, models must be placed in android/app/src/main/assets/ due to TensorFlow Lite's asset loading limitations.
  • For iOS, add models to the Xcode project as before.
  • The code changes to YOLOFileUtils.kt for path prefix stripping are not used in the current loading flow.

• Enhancement: Smart label positioning for detection results
  • Labels for detected objects are now always visible and never cut off at the edges of the screen
  • Implemented boundary checks in all directions (top, bottom, left, right)
  • By default, labels are placed above the box; if off-screen, they are placed inside the box
  • Implemented as a reusable, high-performance function
  • Consistent behavior on both iOS and Android platforms
• Bug Fix: Accurate coordinate handling for all model sizes
  • Fixed bounding box and keypoint placement for all YOLO tasks (detect, segment, pose, OBB) when using models with input sizes other than 640x640

Targeted tasks:

• ✅ Object Detection (DETECT)
• ✅ Instance Segmentation (SEGMENT)
• ✅ Pose Estimation (POSE)
• ✅ Oriented Bounding Box (OBB)
• ⚠️ Classification (CLASSIFY) - Not applicable for label placement

Upgrade note:

• Labels for all detection results are now always visible and neatly placed
• Coordinate handling is robust for all model sizes and tasks
• Example app is more flexible and user-friendly

• Breaking: None - fully backward compatible
• Bug Fix: Fix iOS segmentation mask alignment issue
  • Masks now correctly align with detected objects in both portrait and landscape modes
  • Removed explicit contentsGravity settings that caused mask stretching
  • Simplified mask positioning to match yolo-ios-app reference implementation
• Enhancement: Add mask layer frame update during orientation changes
• Internal: Remove unnecessary margin calculations for mask positioning

• Breaking: None - fully backward compatible
• New Feature: Add frame capture functionality with detection overlays
  • Capture camera frames with bounding boxes, masks, poses, and other overlays
  • Save captured images to device gallery or share with other apps
  • Support for all YOLO tasks (detect, segment, pose, classify, OBB)
  • New captureFrame() method in YOLOViewController returns JPEG image data
• Enhancement: iOS capture includes all overlay types (masks, poses, OBB)
• Enhancement: Android capture with multiple fallback methods for reliability
• Documentation: Added comprehensive frame capture API documentation

• Breaking: None - fully backward compatible
• New Feature: Enable camera preview without valid model path
  • YOLOView now starts with camera-only mode when model is unavailable
  • Graceful error handling instead of crashes on both iOS and Android
• New Feature: Add dynamic model switching via switchModel() method
  • Switch between different models without restarting camera
  • Enables progressive model loading and A/B testing scenarios
• Enhancement: Improved error messages and logging for model loading failures
• Documentation: Added comprehensive examples for new features

• Fix Android landscape orientation coordinate mapping issue
• Add device orientation detection for proper image rotation
• Implement separate image processors for portrait/landscape modes
• Correct aspect ratio calculations for all YOLO tasks in landscape mode

• Add Support for Landscape Mode

• Fixed critical memory leaks in iOS YOLOView disposal and model switching
• Added proper dispose implementation for YOLOView on both iOS and Android platforms
• Fixed native rendering issues for detection visualization
• Fixed Android model label loading issues
• Enhanced single image inference result updates
• Improved resource cleanup when switching between models or tasks

• Merge example READMEs
• Rename example/example.dart to example/main.dart

• Added example/example.dart for usage demonstration.

• Added Dart publish dry run to CI
• Renamed incorrect docs/ directory to /doc

• Added customizable result streaming with YOLOStreamingConfig
  • Enable detailed control based on streaming mode
  • Enable throttling and frame dropping for performance optimization
  • Added optional support for mask and pose data in results
• Added multi-instance YOLO model support
  • Run multiple YOLO models simultaneously
  • Independent configuration for each instance
  • Efficient resource management across instances
• Enhanced Swift backward compatibility
  • Improved support for older iOS versions
  • Better compatibility with legacy Swift code
• Updated documentation
  • Added comprehensive model integration guide
  • Improved API documentation
  • Enhanced troubleshooting section

• Improved publish workflow robustness.

• Fixed publishing workflows for non-sequential version numbers.

• Added example/main.dart for usage demonstration.

• Updated publishing workflows.

• Added example/main.dart for usage demonstration.
• Created shared_main.dart to eliminate duplication between example.dart and main.dart.
• Resolved pub.dev warning: “No example found.”
• Improved pubspec.yaml to explicitly point to the example file.

• Simplified package publishing workflow
• Removed Python-based version check in favor of direct pubspec.yaml version reading
• Improved GitHub Actions workflow reliability
• Fixed tag management and release process

• Add optional confidence and IoU thresholds for single image inference
  • Thresholds can be passed to predict() method for temporary use
  • Does not affect subsequent predictions or camera inference
  • Useful for fine-tuning detection sensitivity per image

• Updated package topics to comply with pub.dev requirements
• Improved package validation and documentation

• Fixed CI/CD pipeline issues for pub.dev publishing

• Updated package validation and documentation
• Improved error handling and logging
• Added support for multiple model types:
  • Object Detection (YOLOv11)
  • Pose Estimation
  • Image Segmentation
  • Oriented Bounding Box (OBB) Detection
  • Image Classification
• Enhanced camera functionality:
  • Camera flipping between front and back cameras
  • Camera zooming with pinch gestures
  • Improved camera preview quality
• Updated package validation and documentation
• Improved error handling and logging
• Added comprehensive example app showcasing all features
• Enhanced documentation with detailed usage examples

• Fixed front camera orientation issue on Android where detection results were displayed upside down.
• Fixed vertical flipping for bounding boxes, segmentation masks, pose keypoints, and OBB (oriented bounding boxes) when using front camera.
• Added proper canvas transformations for segmentation mask rendering with front camera.
• Improved overall detection accuracy and visual alignment for front-facing camera usage.

• Added camera switching functionality to toggle between front and back cameras.
• Added switchCamera() method to YoloViewController for programmatic camera switching.
• Added switchCamera() method to YoloViewState for GlobalKey-based camera switching.
• Updated sample app with camera switching button in the app bar.
• Updated README documentation with examples of camera switching functionality.
• Improved code coverage with additional unit tests.
• Updated codecov badge to show coverage percentage.

• Android: Fixed pose estimation keypoints not displaying correctly by properly implementing object pooling in PoseEstimator.kt.
• Android: Improved segmentation to work with all model classes, not just early ones like "person" and "car".
• Android: Enhanced model metadata loading to extract labels from model files with fallback to COCO dataset classes.
• Android: Fixed lifecycle management in YoloView.kt with proper onLifecycleOwnerAvailable implementation.
• Android: Made Box class fields mutable (var instead of val) to properly support object pooling.
• Performance: Various optimizations for faster inference and more reliable detection.

• iOS: Implemented direct FPS (Frames Per Second) reporting to Flutter, similar to Android. Native-calculated FPS is now included in the data sent to Dart during real-time inference.
• Android: Fixed an issue where the camera preview would remain black by improving native lifecycle management and camera initialization timing. (Previously part of 0.0.9 prep)
• Android: Added detailed debug logs to YoloPlatformView initialization. (Previously part of 0.0.9 prep)
• lib/yolo_view.dart: Added debug logs for communication channel creation and improved null checks. (Previously part of 0.0.9 prep)
• .pubignore: Updated to optimize the content of the published package. (Previously part of 0.0.9 prep)
• General: Incorporated various improvements from previous development versions (including enhanced model path resolution and logging). (Previously part of 0.0.9 prep)

• Android: Fixed an issue where the camera preview would remain black by improving native lifecycle management and camera initialization timing.
• Android: Added detailed debug logs to YoloPlatformView initialization for easier troubleshooting.
• lib/yolo_view.dart: Added debug logs for communication channel creation and improved null checks.
• .pubignore: Updated to optimize the content of the published package.
• General: Incorporated various improvements from previous development versions (including enhanced model path resolution and logging).

• Fix Android implementation for inference results not displaying or updating
• Fix "Unresolved reference: setIoUThreshold" error by fixing method name casing
• Add support for both "setIoUThreshold" and "setIouThreshold" method names for robustness
• Enhance error handling and logging for event channel communication
• Improve StreamHandler implementation for more reliable event dispatching
• Add fallback mechanisms for when direct method calls fail
• Fix reflection-based sink access for CustomStreamHandler
• Add test message mechanism to verify event channel connection
• Significantly increase logging for easier troubleshooting
• Update documentation with clear guidance on model placement and path resolution
• Recommend using model name only (without extension) for best cross-platform compatibility

• Fix iOS implementation for loading .mlmodel files from Flutter assets
• Significantly improve model path resolution for different path formats
• Add extensive logging to help debug model loading issues
• Fix Flutter asset bundle path issues with nested directories

0.0.7 #

• Fix iOS implementation to properly load models from Flutter assets
• Improve asset path resolution for paths like 'assets/models/yolo11n.mlmodel'
• Fix syntax errors in YoloPlugin.swift

• Add iOS implementation for checkModelExists method
• Add iOS implementation for getStoragePaths method
• Fix cross-platform consistency for model path resolution

• Update README to match current implementation of YOLO class constructor
• Fix documentation for threshold management in the API reference
• Add optional controller-based approach for managing YoloView settings
• Make onResult callback truly optional
• Improve threshold controls with IoU threshold support
• Update code documentation with detailed examples
• Add support for direct YoloView state access via GlobalKey
• Enhance error handling and debug logging
• Translate Japanese comments to English

• Initial release
• Object detection with YOLOv8 models
• Segmentation support
• Image classification support
• Pose estimation support
• Oriented Bounding Box (OBB) detection support
• Android/iOS platform support
• Real-time detection with camera feed
• Customizable confidence threshold
• YoloView Flutter widget implementation