DEEP LEARNING WITH JETSON NANO: REAL-TIME OBJECT DETECTION AND RECOGNITION
[1. Інформаційні системи і технології]
Автор: Gura V.T., Department of Radioelectronic and Computer Systems, Faculty of Electronics and Computer Technologies, Ivan Franko National University of L’viv;
Osadchuk O.Y., Department of Radioelectronic and Computer Systems, Faculty of Electronics and Computer Technologies, Ivan Franko National University of L’viv
Jetson Nano is an AI-enabled computer that enables the creation of smart systems. Jetson Nano is a compact solution for AI tasks based on CUDA-X. It delivers 472 gflops of performance in today's AI applications while consuming only 5 watts of power. Jetson Nano supports high-resolution sensors, can process data from multiple sensors simultaneously, and can run multiple neural networks on each sensor stream. It also supports many popular AI frameworks (TensorFlow, PyTorch, Caffe/Caffe2, Keras, MXNet), allowing developers to integrate their preferred models and frameworks into the product.
There are two ways to use this board. It would be interesting to study standard frameworks like Keras and Tensorflow. It will work in principle, but of course it is inferior to a full desktop video card. The task of optimizing the model the user will have to take care of. The second way is to use ready classes provided with the board. It is simpler and works "out of the box", the disadvantage is that the implementation details are much more hidden, in addition, you will have to learn and use custom-sdk.
To begin with, I tried to run model training for image classification directly on the Jetson Nano. In the end the attempt failed - the board overheated in 5 minutes and hung up. For resource-intensive calculations you need a cooler to the board, although by and large there is no point in doing such tasks directly on the Jetson Nano - the model can be trained on the PC, and the finished saved file can be used on the Nano.
The next step was to test pre-trained models for object recognition and object detection and classifications. Some of the most popular models were used to compare the results. Considering that the difference in accuracy between them did not differ more than 1-2 percent, attention was paid to the frame rate per second.
As you can see, NVIDIA's board is quite interesting and quite productive. If someone needs processing power in a compact size, it is well worth it. Even the standard models can be used on Jetson Nano, although with varying success - something works faster, something does not. However, the results can be improved, there are enough instructions to optimize the model and reduce memory size.
1. Jetson Nano: Deep Learning Inference Benchmarks: https://developer.nvidia.com/embedded/jetson-nano-dl-inference-benchmarks