4 ms·
Inference Speed is very important. The TensorFlow object detection model zoo lists timings for their models; we picked our model to meet timing demands on a ce
by dekhn 8y ago
Inference Speed is very important. The TensorFlow object detection model zoo lists timings for their models; we picked our model to meet timing demands on a cell phone.
I don't think CPUs are going to give you decent performance unless you microoptimize the inference for the host architecture and even then the gains are marginal.
If you want to avoid expensive GPUs, take a look at things like Intel's Neural Compute Stick, Google's Vision Kit and the Edge TPU. But you can also run models on modestly priced cell phones- we use $100 cell phones to do object detection and tracking, but it's not pretty.