v0.1.3

Breaking Changes

• RGB everywhere — All image handling now uses RGB channel ordering. BGR conversions happen only at OpenCV boundaries (cv2.imread, cv2.imwrite). If you were passing BGR images directly, convert them first.
• Media class removed — Replaced with VideoReader, CameraStream, and VideoWriter. Use pf.read_video() as a shorthand for VideoReader.
• show_frame → display_video — New display_image and save_image utilities added.
• OCR support removed — from_easyocr, from_tesseract, from_paddleocr, from_ppstructure converters and the ocr_data field on Detection have been removed.
• sha256 parameter removed from assets.download() — Verification is now handled server-side.
• Pillow added as a required dependency.

New Converters

• from_efficienttam() — EfficientTAM and SAM mask outputs
• from_falcon_perception() — Falcon Perception model output with RLE mask decoding
• from_supervision() — Supervision detections
• from_rfdetr() — RF-DETR model output

New Features

• class_names parameter on all detection converters for custom class label mapping
• COCO_CLASSES constant for standard COCO label mapping
• to_pil() / from_pil() image conversion utilities
• DisplayExit exception and quit_key parameter on display_video for controlled loop termination
• assets module with download progress bar, speed display, and cache management
• Adaptive fps_counter overlay in annotators

Bug Fixes

• Deep copy keypoints in Detection.copy() — previously shared references
• Fix metadata fallback to None instead of {} in copy
• All filter methods now return deep-copied detections, preventing mutation of originals

Improvements

• Buffer class refactored to use deque
• Fine-tuned adaptive parameter scaling in annotators
• Simplified docstrings across converters
• Comprehensive test suite added