Hi, I’m retraining FasterRCNN on Openimages and have a couple of queries:
What image size should I be using for the resnet50_fpn backbone in the data loader? (I’ve seen both 128x128 and 256x256 on Kaggle, but the correct value for resnet50 I thought was 224x224?
When calculating IoU how do I know which object in validation matches the inferred target object for any given image? Do I just check all boxes of the given classification and select the best IoU as the correct object?
Any help really appreciated!