Loading...
Please wait, while we are loading the content...
Similar Documents
Embedded Deep Learning for Face Detection and Emotion Recognition with Intel© Movidius (TM) Neural Compute Stick
| Content Provider | Semantic Scholar |
|---|---|
| Author | Nardo, Emanuel Di Petrosino, Alfredo Santopietro, Vincenzo |
| Copyright Year | 2018 |
| Abstract | Nowadays deep learning is one of the main topics in almost every field. It helped 1 to get amazing results in a great number of tasks. The main problem is that this 2 kind of learning and consequently neural networks, that can be defined deep, are 3 resource intensive. They need specialized hardware to perform a computation in a 4 reasonable time. Unfortunately, it is not sufficient to make deep learning "usable" 5 in real life. Many tasks are mandatory to be as much as possible real-time. So it is 6 needed to optimize many components such as code, algorithms, numeric accuracy 7 and hardware, to make them "efficient and usable". All these optimizations can 8 help us to produce incredibly accurate and fast learning models. 9 1 Embedding and face detection 10 Our work focused on two main tasks that have gained significant attention from researchers, that 11 are automated face detection and emotion recognition. Since these are computationally intensive 12 tasks, not much has been specifically developed or optimized for embedded platforms. We show how 13 inference can be accelerated using Intel’s Neural Compute Stick (NCS) [2]. It is a tiny fanless deep 14 learning device, powered by the low power high performance Movidius Myriad 2 Vision Processing 15 Unit (VPU) and allow to accelerate inference optimizing neural networks and operations in order to 16 allow resource-intensive computation on low-resource platforms. We show how this tiny device can 17 let some intensive Deep Learning applications run on embedded devices such as the Raspberry Pi [5] 18 we have used. The inference pipeline is shown in Fig. 1, while Fig. 2 shows how each of the three 19 Neural Compute Stick has been used in the inference acceleration process, two are designed for face 20 detection and the last one to compute the emotion recognition. 21 Figure 1: Raspbery pipeline Submitted to 32nd Conference on Neural Information Processing Systems (NIPS 2018). Do not distribute. For the first task we have used two of the three networks that characterize a MTCNN [6] and applied 22 a non-maxima-suppression for filtering out outliers in order to get a good trade-off between accuracy 23 and performance. 24 Figure 2: Raspbery + NCSs pipeline 2 Emotion recognition 25 Emotion recognition is a very interesting area to deal with on embedded platforms. It extracts the 26 sentiments starting from face movements. Like many other tasks it needs to be near real-time and 27 it is very difficult to get this kind of performance on embedded devices. We show how the NCS 28 acceleration is able to halve SqueezeNet [1] inference time. This kind of network is used before to 29 reduce the total number of parameters in neural networks like AlexNet [3] and to work efficiently on 30 devices that are not able to handle complex neural networks. Test are done on the eNTERFACE [4] 31 dataset that has 6 classes (Fig. 3). 32 (a) Anger (b) Happiness (c) Surprise (d) Fear Figure 3: Example of a subject expressing emotions |
| File Format | PDF HTM / HTML |
| Alternate Webpage(s) | https://openreview.net/pdf?id=SyxkWkdPoX |
| Language | English |
| Access Restriction | Open |
| Content Type | Text |
| Resource Type | Article |
