SimVODIS: Simultaneous Visual Odometry, Object Detection, and Instance Segmentation.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Robotics

Background:

• Intelligent agents require environmental understanding for effective human interaction.
• Current methods often process geometric and semantic information separately, leading to computational inefficiency and architectural complexity.
• A unified approach is needed to integrate environmental perception for advanced agent capabilities.

Purpose of the Study:

• To propose a novel neural architecture, SimVODIS, for simultaneous geometric and semantic environmental perception.
• To overcome the limitations of separate processing methods in terms of computation and software architecture.
• To enhance the autonomy and service capabilities of intelligent agents.

Main Methods:

• Developed SimVODIS, a neural architecture built upon Mask-RCNN.
• Employed supervised training for the core architecture.
• Utilized unlabeled video sequences and photometric consistency for self-supervised training of pose and depth estimation branches.

Main Results:

• SimVODIS achieves state-of-the-art or superior performance in pose estimation, depth map prediction, object detection, and instance segmentation.
• All tasks are executed concurrently within a single computational thread.
• Demonstrated the efficacy of self-supervision for training geometric perception tasks.

Conclusions:

• SimVODIS successfully integrates geometric and semantic environmental understanding in a single framework.
• The proposed architecture offers a computationally efficient and architecturally simpler solution compared to existing methods.
• SimVODIS is expected to significantly advance the autonomy and effectiveness of intelligent agents in real-world applications.