PMIndoor: Pose Rectified Network and Multiple Loss Functions for Self-Supervised Monocular Indoor Depth Estimation.

Area of Science:

• Computer Vision
• Machine Learning
• Robotics

Background:

• Self-supervised monocular depth estimation excels in outdoor scenes but struggles indoors.
• Indoor scenes present unique challenges: non-textured regions and complex, free-hand camera motion.
• Existing methods often fail due to ambiguous photometric losses and inaccurate pose estimation.

Purpose of the Study:

• To propose PMIndoor, a novel self-supervised framework for accurate indoor depth estimation.
• To address the challenges of non-textured regions and complex camera pose in indoor environments.
• To enhance the performance of monocular depth estimation for indoor scenes.

Main Methods:

• PMIndoor utilizes multiple loss functions to constrain depth estimation in non-textured areas.
• A pose rectified network is introduced to accurately estimate rotation between image frames.
• A multi-head self-attention module is incorporated to improve depth estimation accuracy.

Main Results:

• PMIndoor demonstrates superior performance on the NYU Depth V2 indoor dataset.
• The proposed method outperforms previous state-of-the-art techniques for indoor depth estimation.
• Experiments validate the effectiveness of the pose rectified network and multi-head attention.

Conclusions:

• PMIndoor offers a robust solution for self-supervised indoor depth estimation.
• The framework successfully tackles key challenges in indoor scene perception.
• This work advances the capabilities of monocular depth estimation in complex indoor environments.