This is even more true in the Internet sector. The group buys fire to fight the Hundred Regimental War. O2O fire can fully 'open O' and share the economic fire. It is even her boyfriend's girlfriend who can 'share'.
A lot of cottages, and even technology can be a cottage. For example, this year, after the popularity of full-screen and other technologies, AI seems to have become a popular hot spot for mobile phone manufacturers.
Then the question arises. In the end, what are the methods used to implement AI? What will be the result of various methods? Let's start with photography and talk about the three schools of mobile AI photography.
Genre 1: Universal Acceleration for AI Photography with Chips
Last year, Huawei launched the Kirin 970 chip with an NPU. Afterwards, Apple also disclosed the A11 Bionic Neuron. From the perspective of later Huawei, glory-related products, and the Apple iPhoneX, AI-based ASICs were used to implement AI camera functions. , is the core AI capability of this type of mobile phone.
This can be seen as the first school of mobile AI photography: Chip Pie.
From the bottom of the chip to do AI processing capabilities. There are two levels of purpose, the first is to complete the terminal AI calculations in the hardware, you can ensure the real-time camera security and security. After all, photography tasks uploaded to the cloud computing there may be Caton, off After the network can not work, the data uploading server is facing the data is also obvious.
On the other hand, the chip is easily overlooked for the purpose of general acceleration of AI photography tasks. For example, some games require face recognition, some broadcasts read action instructions, and some filters require spatial processing. All in the CPU, GPU on the phone will immediately enter the turtle speed state, can not support the use of cloud processing does not say, but also lost the real-time experience. APP photography needs tens of millions, at present, only the terminal has an AI processing unit to meet These 'unknown needs'.
After all, what exactly is the experience of chip-based AI photography?
Huawei Mate10, Glory V10 and iPhone X capabilities data show that AI photography can be reflected in identification, motion capture, light and shadow analysis, AR and other aspects.
For example, iPhoneX uses the A11 bionic chip, which can be combined with a structured light sensor and a depth gyro to handle tasks such as face and AR, and can handle image recognition quickly.
Next, glory, Huawei's new products, and a new generation of iPhone will inevitably continue this path.
The advantage of this genre is that users can experience the diversification and growth of AI. The mobile phone is not a static AI experience, but it can evolve continuously with the development of ecology and technology. But the problem is that the threshold of chip flow is high. The huge investment in R&D and the waiting period of several years.
Genre II: Based on camera AI
The second AI mobile photography school, can not fail to mention the Google that loves and hates disputes.
As we all know, Google is not betting on hardware technology, but strategically choose strong AI algorithm advantages and cloud computing strengths. This strategy that Google calls AI First has been embodied in Google’s various wearables, home and mobile devices. Above, the pixel series of mobile phones is no exception.
Google Pixel 2, launched last year, gives a very special AI photography mode. It does not have a proprietary AI chip, but it uses algorithms and AI image processing units to complete the compensation of photography capabilities such as dynamic fuzzy photography. Even cut the camera into very The complex imaging unit, to provide the algorithm to achieve depth of field, spatial AI calculations.
This kind of 'obnoxious' approach, probably only Google will do. The core reason behind this is that Google wants users to adapt to all data being uploaded to Google Cloud's lifestyle, and on the other hand, it is also a business that wants to strengthen AI's advantages. Channels.
AI phone photography hidden three major schools, one of which is false AI
The AI secret of Google Mobile Phone is not in the low-level chip. Instead, it hides a dedicated image processing coprocessor, ImageProcessing Unit (IPU), in the camera area. It is specifically designed to combine cloud computing and algorithm clusters to handle AI camera and video tasks. .
But the problem with this is that the image tasks will be handled in specialized parts and rely heavily on cloud computing. But it does indeed bypass the difficulty of developing chips to some extent, making up for the weaknesses in Google's hardware. The idea of letting mobile phone parts AI by themselves depends on Google’s ability to build up the world’s algorithms and cloud computing capabilities. The disadvantages are: In addition to Google, probably no one can try.
Genre 3: Get an APP with algorithms
After the concept of mobile phone AI became hot, domestic mobile phone manufacturers seemed to be irresistible. The "AI photography" message sprang up after the rain, and this gave birth to the third school of AI photography: APP.
The so-called APP is well-understood, thinking about the various beauty shoot cameras we use, dynamic beauty recording applications, etc. These capabilities are basically used today to achieve better results using AI algorithms. Xiu Xiu’s Mito Camera integrates machine learning algorithms in the APP to identify the relationship between the portrait and the background and the light source, thus separating the portraits.
If this algorithm-based 'AI application' or 'AI filter' is directly loaded into the product's camera, this is an AI photography function.
It seems like, Maybe, it is OK?
Not long ago, the red rice Note5 hit the so-called 'thousand yuan AI dual camera' opened the prelude to the propaganda battle on AI. Then vivo X21 also used the AI photography ability as a propaganda. From the product description, the settlement of these two mobile phones can be seen. The solution is basically the 'AI filter' mode mentioned above: Develop a photographic function with certain identification characteristics, and then deploy it in a mobile phone without a card.
The just-released millet Mix2S also took this solution. That is to initialize the algorithm functions such as 'Meitu Camera' in the mobile phone camera. For example, scene recognition photographing, using machine learning for face and human contour recognition, automatic Bokehs and cutouts, etc.
The problem with this kind of mobile phone is that there is no specific unit for handling AI tasks in the chip and camera. Once the high load AI task is running, it needs to call the cloud. The insufficient response speed of the cloud may cause the recognition rate and accuracy of AI photography to decrease. .
For example, when dealing with night shot effects, the mobile phone needs to use AI algorithms such as light source capture, space capture, etc., so as to achieve clear photographing and light source restoration at night.
And this kind of APP will be a bit embarrassing if you want to load AI apps for night shoots. Because of the heavy load of such deep learning, using traditional mobile chips + cloud computing to run such AI shooting tasks, there will be layers of darkness. Faded, unable to focus for a long time. After taking a picture, I have to upload it to the cloud for a long time to deal with. The accuracy and experience are very poor. And once I have no network, it will be even more troublesome. So we can see that most of the current domestic When the mobile phone promotes the so-called AI camera, it does not mention the night shooting, motion capture and other complex tasks.
As a result, the 'AI photographing' in the publicity materials seen by consumers in this genre becomes the entire AI capability of the entire mobile phone. In fact, these capabilities can be achieved by downloading a relevant app on a mobile phone that is not too bad. , It's a little weird to be a main promotional material for a product.