- Remove INPUT_SIZE, NUM_DETECTIONS, NUM_CLASSES constants
- Use generic fallback values (640, 300, 96) instead of named constants
- All model metadata now extracted dynamically from ONNX runtime at initialization
- Fallbacks only used if dynamic extraction fails (safety net)
- Ensures complete consistency between model training and inference
The system now:
1. Extracts input size from ONNX model input tensor shape
2. Extracts detection count from ONNX model output tensor shape
3. Extracts class count from dataset.yaml (via ClassificationManager)
4. Infers output features per detection from model output shape
This eliminates all hardcoded assumptions about the specific model being used.
Related todos: #remove-hardcoded-constants
🤖 Generated with [Claude Code](https://claude.ai/code)
Co-Authored-By: Claude <noreply@anthropic.com>
- Remove ENABLE_MULTI_PREPROCESSING, ENABLE_CONTRAST_ENHANCEMENT, ENABLE_SHARPENING, ENABLE_ULTRALYTICS_PREPROCESSING flags
- Delete detectWithMultiplePreprocessing() method (complex parallel execution)
- Delete enhanceImageForDetection() method (CLAHE contrast enhancement)
- Delete applySharpeningFilter() method (convolution sharpening)
- Delete preprocessOriginalStyle() method (simple resize)
- Simplify detectWithPreprocessing() to only use ultralytics standard preprocessing
- Remove method parameter since only one preprocessing approach is used
- Clean up orphaned comments and documentation
This removes ~150 lines of unused code while maintaining 100% functionality.
Single ultralytics preprocessing provides good accuracy at 640x640 resolution
for small UI elements like pokeballs and shiny icons.
Related todos: #remove-unused-multipass
🤖 Generated with [Claude Code](https://claude.ai/code)
Co-Authored-By: Claude <noreply@anthropic.com>
- Add ClassificationManager singleton to handle class name mappings
- Load class names dynamically from training dataset.yaml (source of truth)
- Remove hardcoded 96-class mapping from YOLOInferenceEngine
- Add SnakeYAML dependency for YAML parsing
- Update all classNames references to use ClassificationManager
- Ensure consistency between model training data and runtime classification
Related todos: #create-classification-module
🤖 Generated with [Claude Code](https://claude.ai/code)
Co-Authored-By: Claude <noreply@anthropic.com>
- Replace hardcoded NUM_DETECTIONS, NUM_CLASSES, NMS_OUTPUT_FEATURES_PER_DETECTION
with runtime extracted values from ONNX model session
- Add extractModelMetadata() method to dynamically determine model properties
- Update parseNMSOutput, matToTensorArray, and postprocessResults to use dynamic variables
- Fallback to constants if runtime extraction fails for safety
- Replace remaining magic numbers with named constants
Related todos: #extract-model-metadata
🤖 Generated with [Claude Code](https://claude.ai/code)
Co-Authored-By: Claude <noreply@anthropic.com>
- Created PGHLog utility with "PGH - " prefix for easy log filtering
- Enhanced MLResult with specific error types (MLErrorType enum)
- Updated MLInferenceEngine interface to use MLResult pattern
- Converted all Log.* calls to PGHLog.* in core components
- Added type-specific error recovery strategies with onErrorType()
- Improved error propagation and debugging capabilities
🤖 Generated with [Claude Code](https://claude.ai/code)
Co-Authored-By: Claude <noreply@anthropic.com>
- Replaced complex color format conversion (20 lines) with ensureBGRFormat() utility
- Simplified RGB conversion logic using ensureRGBFormat() utility
- Reduced preprocessOriginalStyle from 13 lines to 3 lines with safeResize()
- Added comprehensive preprocessing utilities for code reuse
- Improved maintainability and reduced code duplication
🤖 Generated with [Claude Code](https://claude.ai/code)
Co-Authored-By: Claude <noreply@anthropic.com>
Method Decomposition Improvements:
- Extract complex coordinate transformation logic from parseNMSOutput method
- Create TransformedCoordinates data class for type-safe coordinate handling
- Add transformCoordinates() method with clear parameter naming
- Reduce parseNMSOutput method complexity by 60+ lines
Code Quality Benefits:
- Separate coordinate transformation concerns from detection parsing
- Eliminate duplicate transformation logic across different modes
- Improve testability of coordinate transformation independently
- Better readability with focused, single-responsibility methods
Maintainability Impact:
- Coordinate transformation logic now in one place for easier debugging
- Clear separation between transformation logic and detection creation
- Type-safe coordinate handling with data class
- Foundation for unit testing coordinate transformations
🤖 Generated with [Claude Code](https://claude.ai/code)
Co-Authored-By: Claude <noreply@anthropic.com>
Improvements:
- Create YOLOConfig data class with all inference settings consolidated
- Add CoordinateMode and PreprocessingTechnique enums for type safety
- Update YOLOInferenceEngine constructor to accept configurable settings
- Replace scattered constants with config usage in key methods
- Enable runtime configuration of model file, thresholds, and processing options
Maintainability Benefits:
- All YOLO settings in one place for easy adjustment
- Type-safe enums prevent configuration errors
- Default values maintain backward compatibility
- Cleaner separation between configuration and implementation
Performance Impact:
- Thread pool size now configurable via config.threadPoolSize
- Confidence and processing thresholds easily adjustable
- Foundation for A/B testing different configurations
🤖 Generated with [Claude Code](https://claude.ai/code)
Co-Authored-By: Claude <noreply@anthropic.com>
High Priority Fixes:
- Fix thread pool management: Add shared preprocessingExecutor to prevent creating new thread pools per detection
- Add proper thread pool shutdown with grace period in cleanup() method
- Remove memory leak from repeated Executors.newFixedThreadPool() calls
Code Quality Infrastructure:
- Add MatUtils.kt with safe Mat resource management extension functions
- Add MLResult.kt with standardized error handling for ML operations
- Prepare foundation for systematic error handling improvements
Performance Impact:
- Eliminates thread pool creation overhead during inference
- Reduces memory pressure from abandoned thread pools
- Maintains all original ML detection functionality
🤖 Generated with [Claude Code](https://claude.ai/code)
Co-Authored-By: Claude <noreply@anthropic.com>
- Fix DetectionController initialization crash by properly setting up dependency injection
- Update all function signatures to use new MLDetection type instead of legacy Detection
- Remove duplicate class mapping and getClassIdFromName helper method from ScreenCaptureService
- Update EnhancedOCR to work with new MLDetection and BoundingBox types
- Comment out legacy YOLOOnnxDetector references in MainActivity and DetectionController
- Exclude legacy YOLOOnnxDetector.kt from compilation (moved to .bak)
- Fix missing return statement in YOLOInferenceEngine.detectWithMat method
- Add build commands documentation to CLAUDE.md for future reference
- Ensure all preprocessing and detection functionality preserved from original
Architecture Changes:
- ScreenCaptureService now properly uses MLInferenceEngine interface
- DetectionController updated to use MLInferenceEngine instead of legacy detector
- All detection workflows now use unified MLDetection and BoundingBox types
- Maintained 100% backward compatibility of ML detection capabilities
🤖 Generated with [Claude Code](https://claude.ai/code)
Co-Authored-By: Claude <noreply@anthropic.com>
ARCH-002: Extract ML Inference Engine
- Created MLInferenceEngine interface with async detection methods
- Implemented YOLOInferenceEngine preserving ALL YOLOOnnxDetector functionality:
* Complete 96-class mapping with exact original class names
* All preprocessing techniques (ultralytics, enhanced, sharpened, original)
* All coordinate transformation modes (HYBRID, LETTERBOX, DIRECT)
* Weighted NMS and cross-class NMS for semantically related classes
* Confidence mapping and mobile optimization
* Debug features (class filtering, confidence logging)
* Letterbox resize with proper aspect ratio preservation
* CLAHE contrast enhancement and sharpening filters
- Created ImagePreprocessor utility for reusable preprocessing operations
* Configurable preprocessing with letterboxing, normalization, color conversion
* Coordinate transformation utilities for model-to-image space conversion
* Support for different preprocessing configurations
- Updated ScreenCaptureService to use new MLInferenceEngine:
* Replaced YOLOOnnxDetector with MLInferenceEngine dependency injection
* Added class name to class ID mapping for compatibility
* Maintained all existing detection pipeline functionality
* Proper async/await integration with coroutines
- Applied preferred code style throughout:
* Opening braces on new lines for functions and statements
* snake_case for local variables to distinguish from members/parameters
* Consistent formatting matching project standards
- Removed obsolete YOLO implementations (YOLODetector, YOLOTFLiteDetector)
- Preserved all sophisticated ML features: TTA, multi-preprocessing, confidence mapping
🤖 Generated with [Claude Code](https://claude.ai/code)
Co-Authored-By: Claude <noreply@anthropic.com>
Quildra