ultralytics 8.4.208

v8.4.208 - `ultralytics 8.4.208` Release the 0.5 predict NMS IoU default (#27102)

on Python PyPI

🌟 Summary

Ultralytics 8.4.208 ships practical inference, visualization, and performance updates—most notably a 0.5 default NMS IoU for prediction and tracking, while validation keeps its 0.7 default.

📊 Key Changes

• 🎯 More selective prediction and tracking: When iou is not specified, prediction and tracking now use 0.5 instead of 0.7. Validation remains at 0.7, and explicitly provided values still take precedence.
• 🖼️ Better YOLOE visual-prompt results: Calls using visual prompts now default to conf=0.05, helping retain more similar objects whose confidence scores may be low. Other prediction calls are unchanged.
• 🎨 Custom class colors: Class colors can be supplied with dataset metadata and carried through trained or exported models into prediction visualizations.
• ⚡ Performance improvements: Multi-GPU training with gradient accumulation avoids unnecessary gradient synchronization on batches that do not update the optimizer. NMS-free TensorRT 11 FP16 exports can also be up to 3% faster.
• 🌍 Italian documentation access: Italian links were added to documentation, README, and example notebook language bars.
• 📦 Release: Version 8.4.208 publishes these changes; it does not introduce a new model architecture.

🎯 Purpose & Impact

• Prediction and tracking should generally show fewer duplicate overlapping boxes by default. In crowded scenes, some nearby objects may also be suppressed, so users can adjust iou when needed.
• Lowering the visual-prompt confidence threshold can make “find similar objects” workflows more complete, though it may include weaker detections.
• Custom colors make classes easier to distinguish and keep annotation colors consistent across supported visualizations.
• Training and TensorRT speedups can reduce runtime for the affected multi-GPU and export workflows.

What's Changed

• Make Colors the annotation color spec and carry model class colors by @glenn-jocher in #27092
• Speed up NMS-free TensorRT 11 FP16 engines up to 3% by @fcakyon in #27086
• Default YOLOE visual prompts to conf=0.05 by @fcakyon in #27087
• Speed up multi-GPU training with gradient accumulation by @fcakyon in #27091
• Default predict NMS IoU to 0.5, keep validation at 0.7 by @glenn-jocher in #27084
• Add Italiano to the docs language links by @glenn-jocher in #27095
• ultralytics 8.4.208 Release the 0.5 predict NMS IoU default by @glenn-jocher in #27102

Full Changelog: v8.4.207...v8.4.208