First attemption to built licence-plate recognizer. Main task: It has to be light enough to run efficiently on Movidius Intel Compute Stick (10 FPS will be great)
First attempt:
MobilenetV2-SSD
- Depthwise Separable convolutions
- Inverted Residual Block
- Skip Connections and residual blocks
#FIXME
python setup.py install
- GTX 1060 (6gb)
and yes, I am aware that Movidius does not support Pytorch objects. It is also a some kind of challenge for me to get into TF2.
Some really helpful implementations
- Test pretrained mobilenet_v2, as it is. (in progress)
- freeze layer, modified last one
- write some fancy image loader
- pretrain
- test