ultralytics 8.3.215

v8.3.215 - `ultralytics 8.3.215` Segmentation `crop_masks` speedup (#22386)

on Python PyPI

🌟 Summary

Faster, more stable instance segmentation and more reliable exports. This release accelerates mask cropping by ~3x, hardens ONNX/RT-DETR exports, and restores official Jetson JetPack 5 Docker builds — all while refreshing docs and tutorials. ⚡️🛡️🐧

📊 Key Changes

• ✂️ Segmentation speedup (priority)

  • Refactored utils.ops.crop_mask for ~3x faster mask cropping on small batches (e.g., 6 masks on M4 MacBook Pro) by using efficient per-mask slicing for n<50 and vectorized logic for larger batches.
  • Reduced memory use and simplified logic; safer integer box indexing and cleaner interpolation.
  • Additional performance tweak: only filters predictions without masks when needed (amax-based check) to avoid slow indexing on common cases.
  • PR: Segmentation crop_masks speedup (#22386) by @glenn-jocher

• 🧭 ONNX export reliability

  • Sets dynamo=False by default on PyTorch 2.4 to avoid Dynamo-related instability during export; transparent to users.
  • PR: Set dynamo=False by default for ONNX exports (#22388) by @Y-T-G

• 🚫 RT-DETR export safeguards

  • Uses explicit class checks (RTDETRDecoder) and forces nms=False for end-to-end/RT-DETR exports.
  • Enforces ONNX opset defaults (19) and valid range (16–19) for RT-DETR.
  • PR: Prevent nms export for RTDETR models (#22385) by @Y-T-G

• 📦 Export Docker stability

  • Pins torch<=2.8.0 in the Python export Docker image to avoid IMX issues with 2.9.0; keeps numpy==1.26.4.
  • PR: Pin torch<=2.8.0 in Dockerfile-python-export (#22392) by @Laughing-q

• 🐧 Jetson support

  • Re-enables CI builds for NVIDIA Jetson JetPack 5 Docker image (ARM64), restoring official images for edge deployments.
  • PR: Re-enable JetPack5 Docker builds (#22375) by @lakshanthad

• 📚 Docs and examples

  • Updated and added tutorial videos across multiple pages (RT-DETR, YOLO-World, YOLOE, Track mode, Distance Calculation) to improve onboarding.
  • Minor example fix: dataset download path now defaults to dir='datasets'.
  • PR: Docs video refresh (#22408) by @RizwanMunawar

🎯 Purpose & Impact

• Faster segmentation inference and lower memory usage, especially for typical use cases with fewer masks (beneficial across CPU/GPU and Apple Silicon). 🚀
• More robust and predictable ONNX exports on PyTorch 2.4+; fewer surprises in CI/CD and deployments (including Ultralytics HUB). 🛡️
• Correct-by-default RT-DETR exports that avoid unsupported NMS paths and enforce compatible opsets. ✅
• Stable, reproducible Docker-based exports for IMX targets; reduced breakage from upstream library changes. 🔒
• Jetson users get up-to-date, official JetPack 5 images, simplifying ARM64 edge deployments. 🧩
• Clearer learning resources and smoother onboarding with refreshed video tutorials. 🎥

Tip: Upgrade with pip install -U ultralytics and enjoy speedups automatically—no code changes needed for segmentation or export workflows.

What's Changed

• Set dynamo=False by default for ONNX exports by @Y-T-G in #22388
• Pin torch<=2.8.0 in Dockerfile-python-export by @Laughing-q in #22392
• Prevent nms export for RTDETR models by @Y-T-G in #22385
• Re-enable JetPack5 Docker builds on GitHub-Hosted runners by @lakshanthad in #22375
• Add https://youtu.be/leOPZhE0ckg to docs and refresh outdated embeds by @RizwanMunawar in #22408
• ultralytics 8.3.215 Segmentation crop_masks speedup by @glenn-jocher in #22386

Full Changelog: v8.3.214...v8.3.215