Instance Segmentation on LVIS generalization experiment v0.5
20.1mAPMask R-CNN
Evaluation results
| Method | Links | |||||||||
|---|---|---|---|---|---|---|---|---|---|---|
| Mask R-CNNSupervision=Fully Supervised2021.04 | 20.1 | 32.7 | 21.2 | 7.2 | 19.9 | 28.4 | 10.6 | 21.4 | 32.5 | |
| OursSupervision=COCO masks, Backbone initialization=ImageNet pretrained2021.04 | 10.9 | 16 | 11.3 | 8.7 | 10.5 | 17.4 | 9.2 | 12.9 | 14.7 | |
| Ours (rand. init. backbone)Supervision=COCO masks, Backbone initialization=Random2021.04 | 9.6 | 13.9 | 10.4 | 5.1 | 9.2 | 16.8 | 7.5 | 10.7 | 13.9 | |
| ShapeMaskSupervision=COCO masks+LVIS boxes2021.04 | 8.4 | 13.7 | 8.9 | 5.6 | 8.4 | 10.2 | 6.2 | 8.8 | 10.3 | |
| Mask^X R-CNNSupervision=COCO masks+LVIS boxes2021.04 | 5.6 | 9.5 | 5.8 | 2.4 | 5.1 | 7.9 | 3.1 | 5.6 | 7.8 |