You need first install cornernet to obtain the function of corner pooling. I will write an unified installation script and clean the code soon.
- cuda-10.2
- pytorch-1.4.0
- torchvision-0.5.0
- python-3.6.9
All ResNe(x)t based models are trained with 16 images in a mini-batch and frozen batch normalization (i.e., consistent with models in maskrcnn_benchmark).
I re-implement the model and re-train all the models, and thus the results may be a bit different with the ones reported
*I mainly follow the setting of FCOS.
Model | Multi-scale training | FPS | AP (minival) | AP (test) | Link |
---|---|---|---|---|---|
PolarNet_R50_1x | No | - | 39.6 | - | download |
PolarNet_R101_2x | Yes | 6.7 | 43.7 | 44.1 | download |
PolarNet_dcnv2_X101_2x | Yes | 4.3 | 47.3 | 47.8 | download |