Windows GPU Tensorflow Object Detection API and Real-Time Detection
I am testing the new Tensorflow Object Detection API in Python and I have managed to install it on Windows using docker. However, my trainable model (Faster RCNN resnet101 COCO) takes up to 15 seconds to make a prediction (with very good accuracy), perhaps because I'm only using the Tensorflow processor.
My three questions:
- Given the delay, where is the problem? I heard that FAST RCNN was a good model for low latency visual detection, is it due to CPU only execution?
- With such a latency, can you do efficient real-time video processing with tensorflow GPU GPU or use a more popular model like YOLO?
- A popular docker tensorflow GPU use tool is nvidia-docker, but it is not supported on Windows. Should I keep looking for a local prediction solution (or conda), or should I deploy my model directly to a virtual instance using the GPU (I'm comfortable with google cloud platform)?
Any advice and / or good advice regarding live video processing with Tensorflow is greatly appreciated!
source to share
Given the delay, where is the problem? I heard Faster RCNN was a good model for low latency visual detection due to CPU only execution?
Of course, this is because you are using the CPU.
With that latency, can efficient real-time video processing be done using a tensor GPU, or should I use a more popular model like YOLO?
Yolo is fast, but I once used it for the face and the accuracy wasn't that great. But a good alternative.
A popular tensorflow GPU usage tool in docker is nvidia-docker, but is not supported on windows. Should I keep looking for docker (or conda) for local prediction, or should I deploy my model directly to a GPU-powered virtual instance (I'm comfortable with Google Cloud Platform)?
I think you can still use your local GPU on Windows as Tensorflow supports GPUs in python.
And here is an example, just for that. It has a client that can read the webcam or IP camera stream. The server uses the Penson Tensorflow GPU version and is ready to use a pre-trained model for predictions.
Unfortunately Tensoflow does not support tensorflow-serving
in windows. Also, as you said, is Nvidia-Docker
not supported on windows. Bash on windows
does not support GPU. So I think this is the only easy way right now.
source to share