Masked GAN for Unsupervised Depth and Pose Prediction With Scale Consistency.

Area of Science:

• Computer Vision
• Machine Learning
• Robotics

Background:

• Unsupervised monocular depth and visual odometry (VO) estimation commonly use adversarial learning with reconstruction losses.
• Performance is often hindered by occlusions and changing visual fields between frames.

Purpose of the Study:

• To propose a masked generative adversarial network (GAN) for robust unsupervised monocular depth and ego-motion estimation.
• To mitigate the impact of occlusions and visual field variations on estimation accuracy.

Main Methods:

• Introduced MaskNet and a Boolean mask scheme to filter out occluded or changed regions.
• Implemented a scale-consistency loss for accurate long-term camera trajectory estimation.
• Utilized adversarial and geometric image reconstruction losses as primary training signals.

Main Results:

• Demonstrated that each proposed component improves performance.
• Achieved competitive results for both depth and trajectory predictions on KITTI and Make3D datasets.
• The masking strategy effectively reduced negative impacts from occlusions and visual field changes.

Conclusions:

• The masked GAN framework offers a significant advancement in unsupervised monocular depth and ego-motion estimation.
• The proposed methods provide a more accurate and robust solution for monocular sequence analysis.
• This work paves the way for more reliable visual odometry and depth perception in challenging environments.