@alknemeyer Here is the documentation site https://developer.movidius.com/docs link. You can also find this link on the left side. The release notes for NCSDK 2.05 can be found https://movidius.github.io/ncsdk/release_notes.html.
According to the release notes, NCSDK 2.05 supports TensorFlow version 1.7 and the main features that NCSDK 2 brought to the table was the ability to use multiple graphs with one device using fifos (https://movidius.github.io/ncsdk/ncapi/readme.html). All features and functionality from NCSDK 1 should be present with the exception of one sample from the NCAPPZOO that is currently not working.
Although we do have support for some TensorFlow networks out of the box, there are still some TensorFlow ops that aren't supported yet. Keras sits on top of TensorFlow and if it uses some unsupported ops, the model won't be compatible with the current NCSDK.
As for a guide for retraining a pre-trained object detector, we don't have something like that yet.
@alknemeyer Completely understand where you are coming from. Our TensorFlow support is definitely a work in progress. As for Tensorflow obejct detectors, we do have Tiny Yolo v2 on Tensorflow working. There is a simple example at https://github.com/movidius/ncappzoo/tree/ncsdk2/tensorflow/tiny_yolo_v2.