I have implemented Tiny YOLO network v2 on tensorflow and wanted to optimize for Movidius 2.10. I have thought of proceeding with Depth wise Separable convolution or with Group Convolution to reduce parameters and weights.
How far Movidius supports the above operations?
Depth Convolution and Group Convolution are supported by the NCSDK (see NCSDK2 release notes here)
Furthermore, depth wise separable convolution should be supported on the NCSDK2 since the Mobile Net SSD framework uses depth wise separable convolution. So you should be able to implement it in your Tiny YOLO network.
I hope this information is helpful.