I recently implement couple of algorithms for Deep Learning, which provides fast inference on an embedded devices.
One of them is called ‘LCNN’, you can find the original paper here : https://arxiv.org/abs/1611.06473
And I implemented it with tensorflow,
This codes compress alexnet which takes roughly 150ms or more on a single core cpu,
to a sparse convolutional layered network which takes 10~50ms on the same environment.https://github.com/ildoonet/tf-lcnn/raw/master/images/timeline_alexnet.png
So… based on these new technologies, i am looking for an idea to try.
Like openpose on a robot : Human Pose Estimation Deep Learning Model (OpenPose) ROS Package