Russian scientists from the AIRI Institute of Artificial Intelligence have developed a universal machine vision system for three-dimensional object recognition that performs equally well on different test sets. The development can be used in robotics, augmented reality, and 3D scanning.
Previously, it was necessary to create separate models for each dataset, which slowed down the development process. The new system, based on a pure transformer-encoder, simplifies and speeds up this process.
The creation of three-dimensional machine vision systems is limited by small and heterogeneous datasets. The largest dataset contains only about 7,000 scenes, which is several times smaller than the millions of images for generative models.
To solve the problem, Russian scientists created a universal neural network based on a transformer-encoder without optimizations for specific datasets and conducted a large-scale relabeling to reduce the number of object classes.
Experiments have shown that the new model effectively works with a large number of heterogeneous datasets and recognizes objects in different types of "point clouds" from laser radars and three-dimensional scanners. Scientists hope that the development will accelerate the creation of three-dimensional vision systems and improve the quality of their work.
Read also on the topic:
Russian engineers have developed "technical vision" for river and sea vessels (and water drones too)
Alikhanov was shown the first Russian AI robot tractor from Cognitive Pilot
Russian scientists have developed a neuromorphic device for hydrocarbon exploration
