Self-Supervised Monocular Depth Estimation With Self-Perceptual Anomaly Handling.

Area of Science:

• Computer Vision
• Machine Learning
• Robotics

Background:

• Self-supervised learning shows promise for 3-D information extraction from 2-D images.
• Monocular video training is challenged by moving objects and illumination inconsistencies.
• Existing methods struggle with incomplete filtering and implicit disturbances from anomalous pixels.

Purpose of the Study:

• To develop robust monocular depth estimation methods for 3-D scene understanding.
• To address challenges posed by moving objects and the ill-posed nature of monocular depth estimation.
• To improve the accuracy and reliability of 3-D information extraction from monocular videos.

Main Methods:

• A self-reprojection mask was developed to filter moving objects, overcoming illumination inconsistencies.
• A self-statistical mask method was introduced to prevent filtered pixels from implicitly disturbing the reprojection process.
• Self-distillation augmentation consistency loss was employed to mitigate the ill-posed nature of monocular depth estimation.

Main Results:

• The proposed method demonstrates superior performance on the KITTI dataset.
• Performance improvements are particularly notable when evaluating the depth of potential moving objects.
• The techniques effectively filter moving objects and enhance depth estimation robustness.

Conclusions:

• The novel self-supervised learning approach significantly advances 3-D information extraction from monocular videos.
• The proposed masks and loss function effectively address key limitations in existing methods.
• This work contributes to more reliable and accurate 3-D scene reconstruction using single-camera systems.