INV-Flow2PoseNet: Light-Resistant Rigid Object Pose from Optical Flow of RGB-D Images Using Images, Normals and...

Area of Science:

• Computer Vision
• Robotics
• 3D Reconstruction

Background:

• Standard optical flow and pose estimation methods assume constant scene appearance, failing under strong shading changes.
• Dynamic lighting conditions, common in autonomous driving and robotics, violate this brightness constancy assumption.
• Existing techniques struggle with scenarios involving rotating objects or moving light sources.

Purpose of the Study:

• To develop a novel architecture for accurate optical flow and rigid scene transformation estimation.
• To address limitations of traditional methods in environments with significant illumination variations.
• To enable robust 3D reconstruction in challenging, dynamic visual conditions.

Main Methods:

• A new architecture fuses image, vertex, and normal data for illumination-invariant optical flow estimation.
• A coarse-to-fine strategy anchors optical flows globally, mitigating shading-induced errors.
• A secondary architecture predicts rigid transformations using warped vertex and normal maps, with a 3-step procedure for rotational scenarios.

Main Results:

• The method achieves highly accurate optical flows and rigid transformations in difficult scenarios with strong shading changes.
• It demonstrates robustness against illumination variations and significant rotations.
• Evaluated on synthetic and real datasets, including the Kitti Odometry dataset, showing broad applicability.

Conclusions:

• The proposed method offers a robust solution for optical flow and pose estimation under challenging lighting conditions.
• It advances 3D reconstruction capabilities by handling appearance changes effectively.
• The approach provides a valuable tool for applications in robotics, autonomous driving, and computer vision.