Due to the changes between MMDetection v1.x and v2.x, models directly converted from our previous implementation at AttentiveNorm_Detection suffer significant performance drop.
We will train models using this refactored code and release the trained models here.
- All the models are released on Google Drive (download link)
| Architecture | Backbone | Neck | Head | box AP | mask AP | Remarks |
|---|---|---|---|---|---|---|
| ResNet-50 | BN | - | - | 39.2 | 35.4 | MMDet trained |
| AN (w/ BN) | - | - | 40.8 | 36.4 | AttentiveNorm_Detection trained | |
| AN (w/ BN) | GN | AN (w/ GN) | 41.6 | 37.4 | AttentiveNorm_Detection trained | |
| AN (w/ BN) | - | - | 41.5 | 36.8 | This repo | |
| *GN | GN | GN | 40.3 | 35.7 | - | |
| *SN | SN | SN | 41.0 | 36.5 | - |