Efficient and Scalable Object Localization in 3D on Mobile Device.

Area of Science:

• Computer Vision
• Mobile Augmented Reality
• 3D Object Detection

Background:

• 2D object detection lacks real-world context like size and position.
• Existing methods struggle with efficient object classification and localization on mobile devices.
• Mobile offloading techniques are resource-intensive.

Purpose of the Study:

• To propose a novel 3D object localization solution for mobile devices.
• To enable efficient object recognition and real-world coordinate determination.
• To overcome the limitations of 2D object detection on resource-constrained mobile platforms.

Main Methods:

• Combining a 2D object detection Convolutional Neural Network (CNN) with Augmented Reality (AR) technologies.
• Leveraging Simultaneous Localization and Mapping (SLAM) from ARCore for plane detection and camera information.
• Generating 3D cuboid proposals from 2D bounding boxes using AR capabilities.

Main Results:

• The proposed method achieves fast and efficient identification of everyday objects.
• Real-world coordinates of detected objects are accurately determined.
• The approach is optimized for mobile devices with limited resources.

Conclusions:

• The novel approach successfully integrates 2D CNNs with AR for 3D object localization on mobile devices.
• This method provides a robust solution for understanding object size and position in real-world environments.
• The technique is suitable for real-time applications on mobile platforms without requiring significant computational offloading.