NVIDIA develops `` AI to apply animal expressions to other animals '', demo website is also being released
Humans can imagine the face of a “lion yawning like a cat” or “a wolf laughing like a Shiba-inu” in his head, but for the computer, he recognizes the facial expression of the animal, It seems difficult to create an image by applying it to another animal.
[1905.01723] Few-Shot Unsupervised Image-to-Image Translation
AI Uses GANs to Turns Pets into GANimals | NVIDIA Blog
NVIDIA, which develops state-of-the-art GPUs, is actively expanding into the fields of AI and deep learning. In the past, we have developed “software that converts graffiti into landscape photos in real time” and “AI that creates a 3DCG virtual environment from real-world images”.
NVIDIA announces amazing drawing technology `` GauGAN '' that can convert graffiti into real landscape photos in real time-GIGAZINE
NVIDIA reveals how to create a 3DCG virtual world from real video with AI, making it possible to significantly cut the cost of building a 3D environment-gigazine
A new NVIDIA research team has developed an AI that 'recognizes facial expressions from one animal photograph and applies them to the faces of various other animals.' This AI uses an algorithm called a hostile generation network (GAN) that uses two neural networks called a “generation network” and an “identification network” to improve output accuracy.
GAN determines whether the identification network is 'real or fake made by AI' from images generated by the generation network. The generation network learns to deceive the identification network, the identification network learns to more accurately identify what was output by the generation network, and it has a mechanism to improve the accuracy of each .
NVIDIA's new AI created using this algorithm is named “GANimal”. GANimal is easy to run. For example, if you input a photo of a dog that looks like a laugh and surround the face with a square frame ...
Many other types of dogs with similar facial expressions were generated. Looking at the generated image, the dog shows a tongue, but some cats have their mouths closed. However, the tiger in the lower left has a facial expression similar to the input image with the mouth open.
Even with different types of animals, sloth bears , American bears , Rikaon, etc., looked similar to the input images. “Most GAN-based image transformation networks are trained to solve a single task,” says Ming-Yu Liu, a research team, and that a single AI generates multiple images. Claims to be a major advance.
NVIDIA publishes a website that can run a demo of GANimal, and anyone can easily generate images of various animals using photos of pets.
Click 'Browse' from the top page ...
Upload the selected
Click 'Upload' ...
An image is displayed at the bottom of the page. Drag and surround the animal's face with a square frame.
Click 'Translate' ...
Various animal photos were created that looked like the cats in the input image.
Some of the images generated during this demonstration had animals with broken faces, but the research team will continue to improve the accuracy of the tools. They have already tested non-animal subjects and are training on flowers and food images.