Landing AI develops “visual prompting” technology for teaching edge AI
During the Computex 2023 event, Landing AI, a company specializing in cloud-based computer vision solutions, created a visual promotion technology that leverages Nvidia’s Metropolis for Factories. This solution brings the concept of prompt-based interaction to the computer vision ecosystem, expanding on the framework of text prompting used in technologies like ChatGPT.
Just like in ChatGPT, where a text prompt is utilized to express desired text processing tasks, the visual prompting technology enables users to precisely indicate image processing tasks through visual prompts. Leveraging large pre-trained vision transformers, this technology streamlines the development process by requiring only a few straightforward prompts and examples.
Using this approach, users can deploy computer vision systems and make inferences, all within a significantly reduced development time. Incorporating visual prompts alongside the pre-trained vision transformers empowers users to interact efficiently with the system and rapidly achieve their desired outcomes in the field of computer vision.
“Access to Nvidia AI and Metropolis for Factories enables us to enhance and accelerate our Visual Prompting technology and apply it to make spaces and operations safer and more efficient,” says Carl Lewis, senior director of partnership at Landing AI. “Nvidia’s assistance has been invaluable for advancing our technology roadmap, and we are thrilled to accelerate our collaboration through Metropolis.”
What is visual prompting technology?
To streamline the labeling process of images in the training dataset, visual prompting technology leverages large pre-trained vision transformers. Unlike traditional methods that necessitate labeling every image in its entirety, this advanced approach enables users to label only a few specific areas within an image. This efficient labeling strategy accelerates the process, reducing the user’s workload while still achieving accurate and comprehensive training data.
Recognizing the inefficiency of manually labeling each defect in an image, Landing AI says it decided to mimic the way humans teach. Rather than laboriously labeling every single defect in an image, the user tags an area of the image, allowing the machine to autonomously find additional defects.
Landing AI has successfully implemented the visual prompting technology into the workflow of one of its life sciences clients, enabling them to create a deployable cell detection model. The customer faced challenges associated with labeling too many cells, which often led to the omission of some cells in the process. With visual prompting, the client achieved their desired results, reporting that the entire task was completed within 10 minutes.
Article Topics
AI/ML | computer vision | edge AI | Landing AI | model training | Nvidia
Comments