If you already made something, tell us!
[AI] UP Detector Demo
Intel® Apollo Lake in UPSquared + Intel® dev tools look a pretty interesting combination to create deep learning applications that can run on live video.
To demonstrate some of the capabilities, we thought that instead of using the typical pretrained neural network models, it would be fun to train a custom one and 'teach' the UP Squared to detect itself as well as a few of its sisters, classifying a few types of boards and devices (UP, UP Squared and Movidius NCS devices) from live cam video.
At the same time allowed us to get a feeling on the expected performance using GPU vs CPU.
To run the app you need:
- an UP Squared (with Ubuntu 16.04.3).
- USB camera (in our case USB 1080p camera, model AD-BHC7350U, others could work without problems).
- Intel® System Studio software (included in UP Squared AI Vision Development Kit).
- OpenVINO software (included in UP Squared AI Vision Development Kit).
How we do it
We trained a MobileNet SSD V1 neural network, pretrained with coco dataset (available here: https://github.com/tensorflow/models/blob/master/research/object_detection/g3doc/detection_model_zoo.md), to detect up squared, up board and Movidius NCS with over 5000 pictures. The model runs on Tensorflow-gpu 1.4 and supports acceleration by the GPU Intel® Gen 9 HD in the UPSquared.
Once the model was trained, we obtained the inference graph and with model optimizer tools we obtained the xml and bin model.
Attached is the build recipe plus the Intel® System Studio (OpenVINO) packages and source code as well as the network model.