Self-Supervised Monocular Depth Estimation With Positional Shift Depth Variance and Adaptive Disparity Quantization.

Area of Science:

• Computer Vision
• Machine Learning
• 3D Scene Reconstruction

Background:

• Self-supervised learning for 3D scene understanding from monocular videos is challenging due to independently moving objects.
• Existing methods often assume rigid scenes, failing to account for dynamic elements.

Purpose of the Study:

• To develop a robust self-supervised method for Single View Depth Estimation (SVDE) from monocular videos.
• To effectively handle independently moving objects in scene reconstruction tasks.

Main Methods:

• Exploiting pixel positional information to generate 'SPIMO' masks that identify and remove moving objects.
• Introducing an adaptive quantization scheme for improved depth map discretization.
• Utilizing boosting techniques for self-supervised learning of moving object depths.

Main Results:

• Achieved state-of-the-art (SOTA) results on KITTI and CityScapes datasets.
• Demonstrated robustness against moving objects and generalization to high-resolution images.
• Reduced model parameters by four to eight times compared to previous SOTA video-based SVDE methods.

Conclusions:

• Pixel positional information is key to robust SVDE from videos, even with moving objects.
• The proposed SPIMO masks and adaptive quantization significantly enhance depth estimation accuracy.
• This approach offers a more parameter-efficient and effective solution for self-supervised 3D scene understanding from monocular videos.