Estimation of 6D Object Pose Using a 2D Bounding Box.

Area of Science:

• Computer Vision
• Robotics
• Machine Learning

Background:

• Estimating an object's 6-Dimensional (6D) pose from RGB images is a challenging problem in computer vision.
• Current methods often struggle with the inherent ambiguity of quaternion representations (q and -q being equivalent) leading to large L2 losses.
• Accurate 3D pose estimation is crucial for applications like robotic manipulation and autonomous navigation.

Purpose of the Study:

• To develop an efficient method for 6D object pose detection and estimation from single RGB images.
• To address the limitations of existing pose loss functions when using quaternions.
• To improve the accuracy and computational efficiency of 3D pose estimation algorithms.

Main Methods:

• A novel quaternion pose loss function is defined to handle the equivalence of q and -q.
• A new convolutional neural network, Q-Net, is designed for pose estimation, incorporating a normalization layer for unit quaternion outputs.
• The Bounding Box Equation algorithm is proposed for rapid and effective 3D translation estimation from 2D bounding boxes.

Main Results:

• The proposed Q-Net model demonstrates efficient and acceptable performance in 6D object pose estimation.
• The Bounding Box Equation algorithm effectively estimates 3D rotation (R) and translation (t) from a single RGB image.
• Experiments on the LineMod dataset show improvements in terms of L2 loss and computational time compared to existing methods.

Conclusions:

• The developed methodology offers an efficient and accurate solution for 6D object pose estimation from RGB images.
• The novel quaternion pose loss and Bounding Box Equation algorithm can enhance existing 2D detection algorithms into 3D pose estimators.
• This approach provides a viable and computationally efficient alternative for real-world applications requiring precise 3D pose information.