PyTorch Detection Models COCO AP

vision
1

I loaded the PyTorch Detection models as:
model_frcnn = torchvision.models.detection.fasterrcnn_resnet50_fpn(pretrained=True) model_retinanet = torchvision.models.detection.retinanet_resnet50_fpn(pretrained=True)

This should load the models that are pretrained on COCO dataset. I do not do any changes to the models and just evaluate the models on the COCO val2017 dataset using Pycocotools. I get the following results on Faster RCNN and Retinanet:

frrcnn
frrcnn1029×346 8.65 KB

retinanet
retinanet1029×346 8.23 KB

It looks like the AP is 9%. Is this the expected output? How is the AP so low for a pretrained model?