Here lists selected experiment result. The performance is potentially being better if more effort is paid on tuning. See experience.md to communicate training skills.
For training and inference instructions, refer detectron2.md. As the project is keeping upgrading, the pretrained model provided on Google Drive might show better performance compared with the one in table. For more details, please refer to our paper.
| Dataset | Task Method | Quantization method | Model | A/W | Reported | AP | Flags |
|---|---|---|---|---|---|---|---|
| COCO | Retina-Net | - | Torch-18 | 32/32 | - | 31.5 | 1x |
| COCO | Retina-Net | - | Torch-18 | 32/32 | - | 32.8 | 1x, FPN-BN,Head-GN |
| COCO | Retina-Net | - | Torch-18 | 32/32 | - | 33.0 | 1x, FPN-BN,Head-BN |
| COCO | Retina-Net | - | Torch-34 | 32/32 | - | 35.2 | 1x |
| COCO | Retina-Net | - | Torch-50 | 32/32 | - | 36.6 | 1x |
| COCO | Retina-Net | - | Torch-50 | 32/32 | - | 37.8 | 1x, FPN-BN,Head-BN |
| COCO | Retina-Net | - | MSRA-R50 | 32/32 | - | 36.4 | 1x |
| COCO | Retina-Net | - | Torch-18 | 4/4 | - | 34.0 | 1x,Full-BN, Quantize-All |
| COCO | Retina-Net | - | Torch-18 | 3/3 | - | 32.8 | 1x,Full-BN, Quantize-All |
| COCO | Retina-Net | - | Torch-18 | 2/2 | - | 29.6 | 1x,Full-BN, Quantize-All |
| COCO | Retina-Net | - | Torch-34 | 4/4 | - | 37.0 | 1x,Full-BN, Quantize-All |
| COCO | Retina-Net | - | Torch-34 | 3/3 | - | 35.9 | 1x,Full-BN, Quantize-All |
| COCO | Retina-Net | - | Torch-34 | 2/2 | - | 33.1 | 1x,Full-BN, Quantize-All |
| COCO | FCOS | - | MSRA-R50 | 32/32 | - | 38.6 | 1x |
| COCO | FCOS | - | Torch-50 | 32/32 | - | 38.4 | 1x |
| COCO | FCOS | - | Torch-50 | 32/32 | - | 38.5 | 1x,FPN-BN |
| COCO | FCOS | - | Torch-50 | 32/32 | - | 38.9 | 1x,FPN-BN,Head-BN |
| COCO | FCOS | - | Torch-34 | 32/32 | - | 37.3 | 1x |
| COCO | FCOS | - | Torch-18 | 32/32 | - | 32.2 | 1x |
| COCO | FCOS | - | Torch-18 | 32/32 | - | 33.4 | 1x,FPN-BN |
| COCO | FCOS | - | Torch-18 | 32/32 | - | 33.9 | 1x,FPN-BN, FP16 |
| COCO | FCOS | - | Torch-18 | 32/32 | - | 33.9 | 1x,FPN-BN,Head-BN |
| COCO | FCOS | - | Torch-18 | 32/32 | - | 34.3 | 1x,FPN-SyncBN,Head-SyncBN |
| COCO | FCOS | - | Torch-18 | 4/4 | - | 35.2 | 1x,FPN-BN, Quantize-All, double-init |
| COCO | FCOS | - | Torch-18 | 3/3 | - | 34.1 | 1x,FPN-BN, Quantize-All, double-init |
| COCO | FCOS | - | Torch-18 | 2/2 | - | 33.4 | 1x,FPN-BN, Quantize-Backbone, double-init |
| COCO | FCOS | - | Torch-18 | 2/2 | - | 32.0 | 1x,FPN-BN, Quantize-All, singe-pass-init |
| COCO | FCOS | - | Torch-18 | 2/2 | - | 30.3 | 1x,FPN-BN, Quantize-All, double-init |
| COCO | FCOS | LQ-Net | Torch-18 | ter/ter | - | 32.6 | 1x,FPN-BN, Quantize-Backbone, double-init |
| COCO | FCOS | LQ-Net | Torch-18 | ter/ter | - | 26.2 | 1x,FPN-BN, Quantize-All, double-init |
Flags:
FPN-BN indicates adding BN and RELU in the FPN; FP16 implies the case is trained in FP16 (half float) mode; Head-BN represents the prospoal header employes non shared BatchNorm. Full-BN indicates combining FPN-BN and Head-BN. Torch-18/34/50 means the backbone is the Pytorch ResNet-18/34/50.