Introducing Objectron: The Next Phase in 3D Object Understanding
Google AI released its dataset consisting of 15K annotated videos and 4M annotated images.
Most computer vision research has focused on two-dimensional images, and it has reached an exceptional accuracy with the advancements in object prediction.
While 2D object prediction provides bounding boxes, which allows us to know where the object is located in the image, extending this technique for 3D prediction will enable us to capture the object’s size, position, and orientation. Advancing in 3D object prediction has great potential for various applications in robotics, self-driving vehicles, image retrieval, and augmented reality.
Even though 2D object detection methods are mature and have been widely used in the industry, extending these methods for 3D object detection methods from 2D imagery is challenging. This is due to the lack of large real-world datasets of annotated 3D videos compared to 2D images.
Like how ImageNet has enabled computer vision researchers to advance 2D image tasks, the Google AI team has released the Objectron dataset for 3D object detection. This dataset aims to empower the research community to advance 3D object understanding.
