RobotP: A Benchmark Dataset for 6D Object Pose Estimation.

Area of Science:

• Robotics
• Computer Vision
• Machine Learning

Background:

• Deep learning excels in robotic vision, but large datasets for six-dimensional (6D) object pose estimation are scarce due to data collection difficulties.
• Existing methods struggle with representative and sufficiently large training sets, hindering advancements in accurate object localization.

Purpose of the Study:

• To introduce the RobotP dataset, a novel benchmark for 6D object pose estimation.
• To provide a comprehensive resource for training and evaluating learning-based approaches in robotic vision.
• To address the data scarcity challenge in 6D object pose estimation.

Main Methods:

• Utilized a 3D reconstruction pipeline to generate high-quality depth images, ground truth poses, and 3D models of common objects.
• Automated the creation of object segmentation masks and 2D bounding boxes from reconstructed data.
• Synthesized numerous photo-realistic color-and-depth image pairs with ground truth 6D poses to augment the dataset.

Main Results:

• The RobotP dataset, distributed via the Shape Retrieval Challenge, enables unified training and testing of various learning-based pose estimation methods.
• Evaluation revealed significant potential for improvement in 6D object pose estimation, especially for dark-colored objects.
• The use of photo-realistic synthetic data was shown to enhance the performance of pose estimation algorithms.

Conclusions:

• The RobotP dataset provides a valuable resource for advancing 6D object pose estimation research.
• Further research is needed to improve pose estimation accuracy, particularly for challenging object types.
• Synthetic data generation is a promising strategy for overcoming data limitations in robotic vision tasks.