Object Detection on Open Images unseen classes (non-overlapping)
21AR@1000 (Overall)SSD-ag-ad
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| SSD-ag-adBase Architecture=SSD, Class Awareness=Agnostic, Training Protocol=Scratch, Training Method=Adversarial2020.11 | 21 | 2.4 | 21.1 | 31.9 | |
| SSD-agBase Architecture=SSD, Class Awareness=Agnostic, Training Protocol=Scratch2020.11 | 20.6 | 2.3 | 20.6 | 31.4 | |
| SSD-awBase Architecture=SSD, Class Awareness=Aware2020.11 | 19.9 | 1.3 | 14.5 | 35 | |
| SSD-ft-ag-adBase Architecture=SSD, Class Awareness=Agnostic, Training Protocol=Fine-tuned from class-aware baseline, Training Method=Adversarial2020.11 | 19.4 | 1.8 | 18.1 | 30.8 | |
| FRCNN-ag-adBase Architecture=FRCNN, Class Awareness=Agnostic, Training Protocol=Scratch, Training Method=Adversarial2020.11 | 19.2 | 2.8 | 13.1 | 33.7 | |
| SSD-ft-agBase Architecture=SSD, Class Awareness=Agnostic, Training Protocol=Fine-tuned from class-aware baseline2020.11 | 18.8 | 2 | 16.8 | 30.1 | |
| FRCNN-ft-agBase Architecture=FRCNN, Class Awareness=Agnostic, Training Protocol=Fine-tuned from class-aware baseline2020.11 | 18.4 | 2.8 | 13.8 | 31.3 | |
| FRCNN-agBase Architecture=FRCNN, Class Awareness=Agnostic, Training Protocol=Scratch2020.11 | 18.2 | 2.8 | 12.5 | 31.8 | |
| FRCNN-ft-ag-adBase Architecture=FRCNN, Class Awareness=Agnostic, Training Protocol=Fine-tuned from class-aware baseline, Training Method=Adversarial2020.11 | 18.1 | 2.6 | 11.2 | 32.8 | |
| FRCNN-awBase Architecture=FRCNN, Class Awareness=Aware2020.11 | 17 | 1.6 | 8.3 | 33 | |
| FRCNN-aw-propBase Architecture=FRCNN, Class Awareness=Aware, Evaluation Stage=First-stage Region Proposal Network2020.11 | 15.3 | 0.5 | 3.7 | 32.9 |