Unsupervised monocular visual odometry via combining instance and RGB information.

Area of Science:

• Computer Vision
• Robotics
• Artificial Intelligence

Background:

• Unsupervised deep learning has advanced monocular visual odometry (VO).
• Estimating camera ego-motion from monocular RGB images unsupervised remains challenging due to scene complexity.
• Current methods often lack scene understanding, relying solely on raw RGB data.

Purpose of the Study:

• To propose a novel unsupervised monocular VO framework (CI-VO) integrating instance and RGB information.
• To enhance camera ego-motion estimation by incorporating scene understanding.
• To improve the performance of monocular VO in complex real-world scenarios.

Main Methods:

• The proposed CI-VO framework utilizes a two-stage approach.
• Instance maps are generated from monocular images without dataset fine-tuning.
• A fusion feature extraction network processes combined instance and RGB data for pose estimation via CI-PoseNet.

Main Results:

• The CI-VO method demonstrates strong performance on the KITTI odometry and raw datasets.
• The approach effectively estimates camera ego-motion by leveraging combined information.
• Experimental results show superior performance compared to existing mainstream monocular VO methods.

Conclusions:

• Combining instance and RGB information significantly enhances unsupervised monocular VO.
• The proposed CI-VO framework offers a robust solution for camera ego-motion estimation.
• This method advances the capabilities of deep learning in visual odometry tasks.