Hi all — sharing a small HF collection of datasets/benchmarks for **instance segmentation under noisy labels** (COCO-format):
Collection:
What’s inside:
- **VIPER (clean)**: images + clean COCO instance annotations (`instances_{train,val}2017.json`)
- **VIPER-N**: noisy COCO instance annotations for VIPER (annotations-only)
- **COCO-N**: noisy COCO instance annotations for COCO (annotations-only)
All annotations are standard COCO instances JSON, so you can evaluate by swapping the `instances_{train,val}2017.json` paths in your pipeline.
Noise generation / applying to other datasets:
Feedback welcome (especially around dataloader integrations / viewer-friendly previews).