ultralytics 8.4.181

v8.4.181 - ultralytics 8.4.181 Fix `.yml` dataset discovery regression, case-insensitive source lists, non-UTF-8 hashing and `auto_annotate` same-stem labels (#26735)

on Python PyPI

🌟 Summary

Version 8.4.181 focuses on making dataset loading, annotation, tracking, and inference more reliable, with several performance and compatibility improvements. No model architecture changes are included.

📊 Key Changes

• Fixed data-pipeline edge cases: Dataset discovery now prioritizes .yaml files before .yml files, preventing unrelated YAML files from breaking dataset loading. .txt and .csv prediction-source lists work regardless of filename capitalization, and dataset hashing now supports filenames with non-UTF-8 characters.
• Improved automatic annotation: auto_annotate now names label files using the detector’s prediction output path, helping keep labels matched to images with the same filename stem.
• Made rotated-box tracking steadier: Tracked object angles now stay consistent between frames instead of unexpectedly flipping, making angle-based tracking output easier to use.
• Reduced segmentation memory overhead: Native mask processing writes results into a preallocated output stack, avoiding extra full-size copies. A related filtering change also avoids repeated GPU-to-CPU synchronization.
• Fixed small-image depth output: Depth maps for one-row, one-column, and 1×1 images now retain their expected two-dimensional shape.
• Improved training metrics on shared GPU hosts: Platform system metrics report the GPUs used by the training run, rather than including unrelated GPUs on the same machine.
• Updated compatibility checks and guidance: PyTorch–Torchvision checks cover newer releases, JetPack detection includes JetPack 7.2, and TensorRT 11 DLA limitations are clarified. CI also tests newer Python versions with recent PyTorch releases.
• KPTTracker is not part of this release: Its earlier addition was reverted.

🎯 Purpose & Impact

• Dataset setup and prediction workflows should fail less often on common filename and file-format edge cases.
• Tracking, depth prediction, and segmentation results are more consistent, including for unusual image sizes and rotated objects.
• Segmentation may use less temporary memory, while Platform GPU charts more accurately reflect each run.
• Updated compatibility checks and documentation should make it easier to identify supported software and hardware combinations.

What's Changed

• Refresh stale version lookups: torchvision pairing, JetPack 7.2, TensorRT 11 DLA by @glenn-jocher in #26728
• Simplify is_jetson JetPack lookup and run each SlowTests torch shard on its newest Python by @glenn-jocher in #26733
• Report only the training run's GPUs in Platform system metrics by @glenn-jocher in #26734
• Add KPTTracker: keypoint-based tracking for Pose models by @FelixTx in #26530
• Revert "Add KPTTracker: keypoint-based tracking for Pose models (#26530)" by @glenn-jocher in #26739
• Unwrap OBB track angles to the nearest half-turn form by @glenn-jocher in #26729
• Keep (H, W) depth maps for one-row, one-column and 1x1 images by @aswanth-07 in #26740
• Threshold native masks into one preallocated uint8 stack and drop per-mask CUDA syncs by @hameddavodi in #26731
• ultralytics 8.4.181 Fix .yml dataset discovery regression, case-insensitive source lists, non-UTF-8 hashing and auto_annotate same-stem labels by @cainiao33 in #26735

New Contributors

• @hameddavodi made their first contribution in #26731
• @FelixTx made their first contribution in #26530

Full Changelog: v8.4.180...v8.4.181