SPTNet: SuperPoint Tracking Network for Visual SLAM

Area of Science:

• Computer Vision
• Robotics
• Artificial Intelligence

Background:

• Visual Simultaneous Localization and Mapping (SLAM) faces challenges with robustness under large parallax and cumulative drift.
• Existing methods often use isolated local features or optical flow, leading to high computational costs and errors.

Purpose of the Study:

• To develop an efficient multi-task neural network for robust image-to-image correspondence in camera-based SLAM.
• To improve tracking accuracy and reduce computational overhead compared to conventional approaches.

Main Methods:

• Propose SPTNet (SuperPoint Tracking Network), a unified architecture coupling feature detection, description, and dense optical flow prediction.
• Introduce a Hybrid Tracking Module (HTM) with a Predictor-Corrector mechanism using optical flow as a prior and descriptors for correction.
• Employ spatially constrained Sinkhorn optimization to eliminate flow-induced drift and enable efficient feature reuse via a shared backbone.

Main Results:

• SPTNet achieved a false matching rate of 1.8% on HPatches (5-pixel threshold).
• Demonstrated substantial reduction in cumulative drift on long-sequence SLAM benchmarks.
• Maintained a high execution speed of 35 FPS on standard GPUs, indicating a compact footprint.

Conclusions:

• SPTNet offers a robust and efficient solution for image-to-image correspondence in visual SLAM.
• The proposed Predictor-Corrector mechanism effectively addresses limitations of conventional methods.
• SPTNet's performance and speed make it suitable for embedded robotic deployment.