ultralytics 8.4.205

v8.4.205 - `ultralytics 8.4.205` Add scene text OCR solution with YOLO26-Text and PaddleOCR (#27061)

on Python PyPI

🌟 Summary

Release 8.4.205 adds a ready-to-use scene-text OCR solution, led by YOLO26-Text detection paired with PaddleOCR or a custom recognizer, alongside important tiled-inference and export fixes.

📊 Key Changes

• 📝 New OCR solution: Detects text with detection, OBB, or segmentation models, then recognizes it with optional PaddleOCR or a custom callable. Results include recognized text, confidence scores, original-image polygons, detection indices, and reading-order sorting. The core Results API remains unchanged.
• 🔤 YOLO26-Text resources: Adds guidance for the YOLO26-Text checkpoint family and a new OBB training dataset combining COCO-Text and HierText, with 65,325 images and about 1.29 million word polygons.
• 🧩 More reliable tiled inference: Removes covered seam fragments, handles empty tiled predictions on Apple MPS, and fixes related segmentation, validation-memory, and class-count issues.
• 📦 More dependable exports: LiteRT and NXP exports now run on CPU to avoid failures on MPS and incorrect results on CUDA. DEEPX support is updated for DX-COM 2.5.0 and its matching runtime requirements.
• 📱 Expanded deployment guidance: Adds a MediaTek Dimensity Android guide and improves documentation for high-resolution tiling and DEEPX deployment.
• ✅ Small correctness fixes: yolo checks now ignores requirements that do not apply to the current environment, and dataset split fractions reject booleans instead of treating them as numbers.

🎯 Purpose & Impact

• OCR users can build a text-reading workflow from detection through recognition without writing their own crop, ordering, and result-handling pipeline. PaddleOCR remains optional; custom recognizers can be plugged in.
• The new scene-text dataset makes it easier to train word detectors for signs, documents, receipts, and other real-world text. It provides detection annotations, not transcripts for recognition training.
• Tiled predictions should contain fewer duplicate or incomplete boxes at window seams, while fixes improve reliability for tiled segmentation and validation.
• CPU-based LiteRT and NXP export avoids device-specific export problems; DEEPX users should match DX-COM 2.5.0 models with DX-RT 3.5.0 or newer.
• The MediaTek guide clarifies that released Android support covers CPU and GPU inference; its NPU results use a prototype and are not yet supported by the released plugin.

What's Changed

• Add hero, FAQ, and SEO metadata to the High Resolution Tiling guide by @glenn-jocher in #27070
• Add Open in Colab badge to the High Resolution Tiling guide by @glenn-jocher in #27071
• Fix tiled segmentation boxes, mask thresholds, val memory, and class counts by @glenn-jocher in #27073
• Drop seam fragments and fix empty MPS tiles in tiled merges by @glenn-jocher in #27072
• Refresh the tiling guide hero after the seam fix and add the enlarge tip by @glenn-jocher in #27074
• Fix LiteRT and NXP exports that fail on MPS or give wrong outputs on CUDA by @amanharshx in #27075
• Add MediaTek integration documentation for validated LiteRT deployment by @UltralyticsAssistant in #27079
• Skip requirements whose environment markers do not apply in yolo checks by @MohammadHijjawi97 in #27078
• Reject boolean split fraction values by @nethmitharushika56 in #27068
• Update DEEPX export to DX-COM 2.5.0 and DX-RT 3.5.0 runtime docs by @deepx-dgkim in #27077
• Add Scene Text OBB dataset by @Laughing-q in #27076
• ultralytics 8.4.205 Add scene text OCR solution with YOLO26-Text and PaddleOCR by @Laughing-q in #27061

New Contributors

• @nethmitharushika56 made their first contribution in #27068

Full Changelog: v8.4.204...v8.4.205