We are using two models, Faster-RCNN and Mask-RCNN. The models configurations are available here and here.
We are using the Detectron2
library for our experiments. Using the weights file and the COCOEvaluator class of
the detectron2, one can replicate the results of the paper.
| Name | Attribute | box mAP | download |
|---|---|---|---|
| real | aircraft | 73.32 | model |
| synthetic | aircraft | 54.86 | model |
| fine_tuning | aircraft | 69.16 | model |
| real | role | 68.21 | model |
| synthetic | role | 35.88 | model |
| fine_tuning | role | 63.10 | model |
| Name | Attribute | mask mAP | download |
|---|---|---|---|
| real | aircraft | 73.67 | model |
| synthetic | aircraft | 56.28 | model |
| fine_tuning | aircraft | 70.51 | model |
| real | role | 69.57 | model |
| synthetic | role | 39.46 | model |
| fine_tuning | role | 67.11 | model |