A Variable Photo-Model Method for Object Pose and Size Estimation with Stereo Vision in a Complex Home Scene

Area of Science:

• Robotics and Computer Vision
• Artificial Intelligence

Background:

• Model-based stereo vision methods are crucial for estimating 6D object poses, enabling robots to perform grasping tasks in complex environments.
• Current methods often require extensive data or pre-training for specific object recognition.

Purpose of the Study:

• To present a novel variable photo-model method for estimating the pose and size of unknown rigid objects using a single photograph.
• To offer a versatile alternative to data-driven approaches for robotic grasping applications.

Main Methods:

• Utilizes a pre-trained You Only Look Once (YOLO) v4 weight for object detection and 2D model generation from a single photo.
• Converts the 2D photo-model into 3D flat photo-models of varying sizes and poses.
• Employs perspective projection and model matching, optimized by a genetic algorithm (GA), to find the best object match in stereo images.

Main Results:

• The variable photo-model method effectively estimates the pose and size of target objects within the same class using single-photo input.
• Indoor experiments validate the method's performance in object pose and size estimation.

Conclusions:

• The variable photo-model method provides a versatile and data-efficient solution for object pose and size estimation.
• This approach has practical implications for robotic grasping, simplifying object detection with limited data requirements.