different inference performance

[wiamadaya]

i have a RetinaNet model trained on a custom dataset, however the inference/prediction is not identical if i run the model on the same image - from your application compared to inference notebook (https://github.com/fizyr/keras-retinanet/blob/master/examples/ResNet50RetinaNet.ipynb) - i have set the same score threshold for both (0.3)

[virajmavani]

Hi @wiamadaya!

The issue here can be the fact that Anno-Mage resizes the images to a max of 500 width or height as in the Pascal VOC dataset. Due to this resizing and loss of information thereafter, there may be some misclassifications.

I understand that custom sized image annotations make a lot of sense and it is on the agenda for further development of the project.

Thank you!