metadata
library_name: transformers
pipeline_tag: image-segmentation
tags:
- instance-segmentation
- mask2former
BOOST
Mask2Former (facebook/mask2former-swin-large-coco-instance) fine-tuned for instance segmentation of: booster, spat.
Trained on full-resolution tiles of 1024 px at image scale 1. Use sliding-window inference over the whole image (see deploy_config.json); running the model on a downscaled full image will miss small objects.
Test results (held-out split)
| mask AP | AP50 | AP75 | precision | recall | F1 |
|---|---|---|---|---|---|
| 0.458 | 0.643 | 0.509 | 0.719 | 0.839 | 0.774 |
Precision, recall and F1 at score >= 0.95, IoU >= 0.5.