I'm currently implementing an Android app for street sign recognition. My solution works quite well for the GTSRB dataset, since it provides a labeled test set of centered images. However, it doesn't scale up to more realistic scenarios like for images in the GTSDB, where the signs only take up some pixels. Is it still recommended to downsample the image to 224x224?
There is a big difference between the 2 datasets. When you train on the GTSRB dataset, your model performs image recognition, while for the GTSDB dataset you need to perform object detection. These are 2 related but different tasks (with detection being considerably harder).
You can consider using a sliding window approach, or use some prior attention model that will learn to extract the area of the signs for later classification.
I don't think that it is recommended to down sample images of 1360x800 pixels to 224x224. (224x224)/(1360x800)=0.046. You are loosing 96% of the information (including your street signs which might reduced to nothing), granted most of it is not interesting, but simple down-sampling doesn't get to choose what to forfeit.