AI-based high-resolution depth sensor for close imaging sets new records

Kyocera Corporation introduces a high-resolution AI-based depth sensor that is utilised to measure tiny objects that have been difficult to measure, utilising conventional depth-sensing technologies. Its new camera delivers record-setting depth measurements with 100μm resolution at a 10cm range, even from reflective or semi-transparent objects. The company's innovation will support manufacturing, medicine, logistics, and various other fields that need automated identification and precise depth measurement, unlocking the potential of AI and robotics with vision capabilities far exceeding the human eye.

A unique configuration including two lenses on a single sensor, the company's AI-based depth sensor delivers the industry's highest-resolution depth measurements to date among stereo cameras. The extremely narrow baseline of its depth sensor enables it to calculate the positional disparities of an object through the left and right lenses at a shorter distance than conventional methods. This precision allows the sensor to make accurate measurements of even the smallest objects.

The AI stereo vision algorithms allow precise depth measurements of reflective or semi-transparent objects. Traditional stereo vision algorithms involve matching objects between left and right images. However, reflective or semi-transparent objects often lack the required contrast or are challenging to identify as the same object, causing measurement errors. The company's AI-based methods employ extensive training data to measure accurately, even with challenging reflective or semi-transparent objects.

Together with their advantages, AI-based approaches often include high annotation costs and extended training times due to the extensive data required for high accuracy. The company, therefore, developed two key technologies to reduce training costs: label-free pre-training and data generation using computing graphics for ten times more precise measurements.

The company's distinctive AI solution uses pre-training technology without labels, supplying equivalent recognition with only 10% of the training data. To manage the issue of conventional AI needing large amounts of training data, it developed a computer graphic data-generation technology. This technology allows the automatic generation of training data in a CG simulation environment that accurately reproduces the target objects and settings. Also, it improved and accelerated the CG rendering calculation method. This CG simulation for AI training makes it feasible to adapt to new objects and environments, allowing for highly accurate 3D distance measurement, even with reflective or semi-transparent objects.