I have a custom dataet and using de COCO API to evaluate
1- Faster RCNN with resnet 50 FPN using the official implementation from torchvision
2- Faster RCNN with resnet 50 FPN using the official Detectron2
The results should be similar but i found to D2 like 3-4 points over torchvision.
In both i desactivated all data augmentation and normalized the input size of my images to 1333x800
My question is, what has D2 that make it better than Torchvision in the AP?
PDT: i am not using the “fast coco implementation” of D2 but the official