Let me answer this question in general term rather than particular specific to any SSD implementation. For many object detection software, there is a pre-processing step when preparing the data. Some kind of cropping and/or rescaling can be done to change the resolution to one supported. The exact algorithm depends on the exact implementation of how the model is originally trained. For some package, that is automatically done when you configured the input image size in your configuration file. So read the documentation or the code sometimes to see what is the specific required way to do it.