Robust Model Fitting via Motion-Aware Pyramid Transformer-Guided Preference Filtering and Consensus Smoothing

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning

Background:

• Robust model fitting is crucial for estimating parameters from noisy data.
• Traditional methods like RANSAC are limited by hypothesis ambiguity and inefficiency.
• Existing learning-based methods lack motion cues for dynamic scenes and global context capture.

Purpose of the Study:

• To propose MPCFormer, a motion-aware Transformer for robust model fitting.
• To address limitations of existing methods in handling dynamic scenes and global context.
• To improve accuracy and efficiency in model fitting with high outlier ratios.

Main Methods:

• MPCFormer integrates correspondence learning with spatiotemporal motion cues, eliminating iterative sampling.
• A motion preference filter explores multi-channel motion information using residual-connected Transformer layers and multi-head preference attention.
• A pyramid consensus smoother with multi-scale Transformer encoding captures local-to-global motion consistency.

Main Results:

• MPCFormer achieves superior performance compared to state-of-the-art methods.
• Demonstrated improvements include 4.68% mAP@5°, 1.89% AUC@3 pixel, and 1.52% F-score.
• The method remains effective even at extreme outlier ratios (up to 95%).

Conclusions:

• MPCFormer offers a robust and efficient solution for model fitting in computer vision.
• The integration of motion awareness and multi-scale context significantly enhances performance.
• This approach provides a strong foundation for analyzing complex dynamic scenes.