From my perspective, having used/programmed an OAK-D for a few months and researched things like TinyML, TFLite, etc., I think it comes down to if you have an ML pipeline/framework you like vs. others.
They're all trying to get in as the device you stick at the head of your image/vision-processing pipeline(i.e., on your robot), but Coral vs. OAK-D is really about TensorFlow and Google's TPU vs. OpenVINO and Intel's
Movidius platform. There are tools in both systems that will let you take a larger, pre-trained CNN like ImageNet or YOLO and convert it into something that will run on on of these smaller platforms and if you're happy with
pre-trained 'nets, that's fine. If, on the other hand, you have a lot of time & tooling invested in Tensorflow(for example), it's going to be a lot easier to run your TF model on a TPU-based board like the Coral.
HTH,
'dillo