Open richard-chau opened 5 years ago
Sounds interesting, will have a try when I got time
Recently trained a new model with mobilenet_v2 as encoder. Havent quite tested it yet, probably you need to double check if it can run in real-time in your case. gcn2_mobile2_320x240.zip Performance wise, the model can handle larger rotation than the previous models in my preliminary tests.
Thanks! BTW, could you also share the network.py ?
Updated in Network.md. Besides, the mobilenet will not help to improve the speed since both GCNv2 and GCNv2_tiny are smaller than it and without batch normalization (could be good for performance though).
Does the author try the Mobilenet design? How much speedup and will performance change much?