π Summary
YOLO 8.4.215 improves segmentation memory efficiency and fixes several data, tracking, and reliability issuesβwithout changing model architectures or default prediction settings.
π Key Changes
- π§ Lower-memory segmentation workflows (main release change): Plotting, saving, displaying, moving, and prompting segmentation masks can now work with masks stored inside their bounding boxes instead of building a full image-sized mask stack. Outputs and defaults remain unchanged; requesting
masks.datastill creates the dense masks. - π·οΈ More reliable automatic annotations: Class labels stay matched to their detections when SAM filters out some masks.
- π₯ More complete video predictions: Video loading now reads frames until decoding ends, rather than stopping at a possibly inaccurate frame count in the file metadata.
- π§ More dependable tracking fallback: If ECC camera-motion estimation fails, the tracker now retains the identity transform instead of potentially using a partially modified estimate.
- π Accurate checkpoint provenance: Checkpoints record the Git commit present when the run starts, even if the checkout changes later.
- π§ Updated Brain Tumor dataset: Now contains 1,115 brain MRI images, one
tumorclass, and a patient-based validation split. Documentation clarifies that it is for research and education, not clinical diagnosis. - π οΈ Improved CI resilience and documentation: CI retries a failed asset-cache restore once, and several documentation page titles are clearer and more distinct.
π― Purpose & Impact
- π Segmentation workflows can avoid the memory and processing cost of a dense stack when they only need masks within detected boxes. This can help with larger images or many detections while preserving existing outputs.
- β Annotation, video, and tracking fixes reduce the chance of incorrect labels, silently skipped frames, or unreliable camera-motion compensation.
- π Checkpoint metadata makes it easier to identify the exact code used for a training run.
- π The revised dataset split and guidance make the Brain Tumor dataset easier to use appropriately, while its limitations remain important: it has no tumor-free examples and is not suitable for clinical diagnosis.
- π§ No new model architecture or change to prediction defaults is introduced in this release.
What's Changed
- Record the launch commit in checkpoint git metadata by @fcakyon in #27213
- Fix ECC identity fallback after failed motion estimation by @AhmetMert01 in #27219
- Give the settings docs page a distinct title by @miles-deans-ultralytics in #26595
- Keep auto-annotation classes aligned when SAM drops masks by @aswanth-07 in #27214
- Shorten front-matter titles on 25 docs pages by @miles-deans-ultralytics in #26316
- Read video frames until capture ends by @Vaishnavi220506 in #27215
- fix: retry failed CI asset restores once by @fcakyon in #27220
- Update brain-tumor dataset to one tumor class with a patient-id validation split by @raimbekovm in #27218
ultralytics 8.4.215Plot, convert and prompt segmentation masks without the dense mask stack by @glenn-jocher in #27216
Full Changelog: v8.4.214...v8.4.215