ultralytics 8.4.212

v8.4.212 - `ultralytics 8.4.212` Read Korean, Cyrillic, Arabic and other scripts with OCR (#27167)

on Python PyPI

🌟 Summary

Ultralytics 8.4.212 expands OCR to more writing systems and improves Platform discovery, data import, training reliability, and inference compatibility.

📊 Key Changes

• 🌍 OCR supports more languages and scripts. Choose a PP-OCRv5 model to recognize Korean, Cyrillic, Arabic, Devanagari, Thai, Greek, Tamil, or Telugu text. These models run through ONNX Runtime, without requiring PaddlePaddle or PaddleOCR. For example:

  ocr = solutions.OCR(recognizer="korean_PP-OCRv5_mobile_rec")
  result = ocr(im0)

  Select the model for the text you expect; OCR does not automatically identify or switch languages. Hebrew is not included in the new models.
• 🔎 Platform Explore adds a Models tab. Search public models by what they detect, browse results, and try or clone models. The Platform docs also cover direct imports from hosted Label Studio, including supported detection, oriented-box, segmentation, and classification projects.
• 🧰 Training and evaluation are more dependable. Fixes prevent distributed training from hanging when canceled, restore deterministic training for certain tasks on older macOS MPS setups, and include mask, pose, and semantic pixel-accuracy details in exported validation summaries.
• ⚡ Inference and export fixes improve compatibility and efficiency. ONNX Runtime is better matched to CUDA 12 PyTorch installs and uses more suitable CPU thread limits; TensorRT checks recognize installations in NVIDIA containers and JetPack. MNN inputs, YOLOE tracking with ReID, and Edge TPU pose keypoints also receive fixes.
• 🛡️ Data handling is safer and more efficient. Detection crop saving avoids unnecessary full-image copies; segmentation-mask conversion now rejects filename collisions before overwriting labels; source lists consistently use UTF-8.
• 📚 Documentation is refreshed. Guides reflect current Platform features, and outdated DeepSparse documentation is removed because the project has reached end of life.

🎯 Purpose & Impact

• People working with multilingual images can use a suitable OCR recognizer for scripts that the default PP-OCRv6 model does not cover.
• Platform users can more easily find relevant public models and bring supported hosted Label Studio projects into their workflow.
• Training, validation exports, and deployment should behave more reliably across distributed setups and a wider range of hardware and software versions.
• Safer file handling and reduced image copying help prevent accidental data loss and avoid unnecessary work.

What's Changed

• Platform docs: Explore Models tab, HEIC previews in Predict, Recent Activity dataset status by @pderrenger in #27160
• Platform docs: current VLM leaderboard picks, Explore model deploy and ranking wording, Predict screenshots without Ask AI by @glenn-jocher in #27163
• Fix DDP hang when Platform cancels training mid-epoch by @artest08 in #27161
• Fix deterministic semantic, depth and segment training on MPS with torch<2.3 by @raimbekovm in #27159
• Divide the dataloader worker cap by the GPUs of the run by @raimbekovm in #27162
• Update YouTube video link in isolating-segmentation-objects docs by @RizwanMunawar in #27165
• Remove neural magic deepsparse docs pages after end of life by @raimbekovm in #27169
• Avoid full-frame copies when saving detection crops by @aswanth-07 in #27170
• Fix YOLOE tracking with native ReID features by @raimbekovm in #27171
• Read source manifests as UTF-8 by @mh-masum6875 in #27172
• Prevent mask conversion from overwriting colliding label files by @AhmetMert01 in #27174
• Platform docs: a clone goes to another workspace; duplicate in place with Copy and Paste by @glenn-jocher in #27176
• Platform docs: direct Label Studio project import, image delete confirmation by @pderrenger in #27178
• Platform docs: Label Studio and Roboflow import screenshots show the real storage limit by @glenn-jocher in #27180
• Fix misplaced Edge TPU pose keypoints for models with x, y-only keypoints like tiger-pose by @raimbekovm in #27181
• Skip a redundant 3 GB TensorRT download on JetPack 7, DGX and NVIDIA containers by @glenn-jocher in #27185
• Fix MNN models returning wrong detections with device=mps or permuted tensor inputs by @raimbekovm in #27177
• Add per-class mask mAP, pose mAP and semantic pixel accuracy to validation CSV and JSON exports by @raimbekovm in #27184
• Keep ONNX models running on the GPU with CUDA 12 PyTorch by @Zephyr4772 in #27175
• Faster CPU ONNX predictions on hyperthreaded machines and CPU-limited containers by @var1914 in #27182
• ultralytics 8.4.212 Read Korean, Cyrillic, Arabic and other scripts with OCR by @Laughing-q in #27167

New Contributors

• @AhmetMert01 made their first contribution in #27174

Full Changelog: v8.4.211...v8.4.212