FocalPose++: Focal Length and Object Pose Estimation via Render and Compare.

Area of Science:

• Computer Vision
• Machine Learning
• Robotics

Background:

• Accurate 6D pose estimation is crucial for robotics and augmented reality.
• Estimating camera focal length alongside object pose is challenging but important for precise scene reconstruction.

Purpose of the Study:

• To develop a novel method, FocalPose++, for joint estimation of camera-object 6D pose and focal length from single RGB images.
• To improve upon existing state-of-the-art render-and-compare methods for pose estimation.

Main Methods:

• Introduced a focal length update rule extending existing render-and-compare 6D pose estimators.
• Investigated and combined direct focal length regression with a disentangled reprojection loss.
• Explored the impact of synthetic training data distributions, finding that parametric distributions fitted on real data perform best.

Main Results:

• Achieved lower error in both focal length and 6D pose estimation compared to state-of-the-art methods.
• Demonstrated improved performance through a combination of novel loss functions and optimized synthetic training data.
• Validated the method on three challenging benchmark datasets with known 3D models in uncontrolled environments.

Conclusions:

• FocalPose++ offers a robust solution for joint 6D pose and focal length estimation.
• The proposed method enhances accuracy and robustness in real-world scenarios.
• Optimized synthetic data generation is key to achieving superior performance in camera-pose estimation tasks.