YOLOv5 Segmentation Dataloader Updates

[glenn-jocher]

Adds automatic exploitation of segmentation labels, if available.

YOLO segmentation label format: class, xy1, xy2, ...

🛠️ PR Summary

_{Made with ❤️ by Ultralytics Actions}

🌟 Summary

Enhanced COCO dataset handling and added support for segmentation data in YOLOv5.

📊 Key Changes

• The get_coco.sh script now supports downloading coco2017labels-segments.zip for segmentation data.
• Updates to datasets.py to handle segmentations and cache this additional data, along with improvements to caching logic and display.
• Extended random_perspective() in datasets.py to handle segments alongside targets, allowing augmented training with segmentation data.
• Added new utility functions in general.py for working with segment-to-box conversions and resampling.

🎯 Purpose & Impact

• These improvements aim to facilitate training on datasets containing segmentation information, a step beyond standard bounding box labels.
• Users can explore more advanced object detection models that incorporate shape information from segmentation masks.
• This update also streamlines caching operations to improve the efficiency of dataset preparation for training, which can lead to speed improvements and better handling of large or complex datasets.