Yang Jing Published from Aofei Temple
Quantum bits Report | Official Account QbitAI
Automatic driving, image generation, 2D to 3D...
Which AI papers are the most popular in 2021? Which papers are the most breakthrough?
Recently, an ÉTS graduate student and YouTube blogger summarized the latest AI breakthrough list this year, with videos, articles and codes available.
We sorted out eight categories from it, and let’s take a look with you Kang Kang~
Video bloggers read it and
The ones that sorted out the most are some technologies that are beneficial to video bloggers.
For example, this TimeLens can create slow-motion videos, with a maximum range that can be expanded from the original 30 frames to 900 frames.
For example, this editing artifact VGPNN - a single video is generated in a variety of seconds.
functions such as deleting or adding someone, changing the background, elongating the time, changing the aspect ratio, and resolution are all basic operations in front of it.
also uses AI to separate the image quality processing, such as moving objects in the image without affecting the background or other objects; using AI to separate the sounds, voice, music and sound effects in the real world...
specifically uses the image quality processing this year. This year, Intel Intel used NVIDIA graphics card to make image quality enhancement patches. In June this year, this demo became popular across the Internet.In order to make the effect more realistic on GTA, the researchers changed the three characteristics in the video: increasing the luster of the car, improving the overall appearance of the vegetation, and making the asphalt pavement look smoother.
In response to this, some netizens said that this is much cheaper than path tracking.
If the raging epidemic has made video conferencing popular, then video conferencing software has brought background replacement technology to the forefront.
Google researchers proposed a method of re-illumination Total Relighting to replace the portraits in the background.
It can re-lit any portrait based on the newly added scene light to look more realistic. The method of
can be further extended to movies and professional video production, and can be used by up owners.
In addition, in addition to background replacement, there is also text replacement, and the style is still retained.
This year, Facebook proposed an AI model that can directly translate or edit text in images and follow the same style.
is similar to this~
DALL·E led image generation
image generation field, the most breakthrough is DALL·E——OpenAI "AI Designer" launched in the New Year, Ng likes.
simply put it into consideration, put forward your text requirements, and it generates images. In principle, it is similar to the extended version of GPT-3 in the direction of text synthesis images.
For example, enter "OpenAI Company Facade", and it can give you more than a dozen design drawings for you to choose from.
There are also progress such as generating images based on hand-drawn sketches, using random differential equations for image synthesis and editing.
2D image generation 3D model
This is another highly popular research direction in the field of AI in 2021.
Just imagine how cool it would be if you only take a photo of an object in real life, you can create a 3D format to insert it into a video or game.
ShaRF proposed by Google Research Institute can be done, such as taking a random chair.
Nvidia also proposed a similar solution GANverse3D, and can create customizable 3D animations with just one image.
and fake 3D scenes that were popular on the outside network some time ago were also rendered through a set of photos.
and LASR model - single out an object from short videos to create a 3D model of humans or animals... There are many similar methods.
Everything can be combined with Transformer
Have you ever thought about combining CNN with Transformer?
In 2021, "cross-border output" sets off a trend in the field of AI.
Based on CNN efficiency and Transformer's expression ability, researchers at Hotelberg University in Germany proposed a method for high-resolution image generation - Tl;DR.
is not just CNN and Transformer.
Stanford and Facebook researchers proposed GANsformers - based on the attention mechanism of Transformer in the StyleGAN2 architecture to generate scene pictures.
Application layer: fitting room, weather forecast
In addition, there is also an extension of the application level based on the original model.
Just like Google proposed an improved version based on the StyleGAN2 architecture, creating an AI online fitting room.
only needs to provide one image of your to automatically try on any clothes.
and researchers like University of Barcelona have developed a deep learning-based approach that can automatically detect floating garbage from aerial images and calculate the amount.
For this purpose, they also created an APP where users can recognize the garbage in sea surface images.
and Apple also proposes an ML algorithm applied to photo albums, and automatically recognizes people in private photos on iOS 15; DeepMind proposes a radar depth generation model to predict weather more accurately.
AI-powered Cyberpunk arm
Researchers from the University of Minnesota have created a Cyberpunk arm - an AI-powered neural interface.
According to reports, amputees can control their arms as dexterously as ordinary people.
programming artifact: GitHub Copilot
For developers, the most breakthrough progress this year is the programming artifact - GitHub Copilot, jointly developed by GitHub and openAI.
just describe the command you want to execute and generate the corresponding code.
or even programmers can write a comment, and Github Copilot can complete the remaining code and make suggestions for improvement, saving programmers a lot of time to search.
Tesla 's autonomous driving
worth mentioning that Tesla's autonomous driving is also selected this time.
On Tesla's AI day, Andrej Karpathy, director of artificial intelligence , showed how Tesla can obtain images to road navigation through 8 cameras.
This includes operations such as compressing data, two-dimensional conversion to three-dimensional output.
...
In addition, in the face of the third wave of artificial intelligence, researchers think about fast and slow in AI; AI forges personal files similar to "Tantan" to discuss whether humans will slide to the right; how Transformer replaced CNN in the CV field? Interested friends who are interested in
can click the link below to learn more details~
is still being updated.
GitHub link:
https://github.com/louisfb01/best_AI_papers_2021
—End —
Quantum bitsQbitAI · Toutiao Sign
Follow us and learn about cutting-edge technology dynamics
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-