Evaluating Image Matching With Robust Estimators: Bridging Natural and Surgical Domains to Enhance Scene...

Area of Science:

• Computer Vision
• Medical Imaging
• Surgical Technology

Background:

• Image matching methods generalize well in natural scenes but struggle in complex surgical environments due to factors like poor lighting and occlusions.
• Keypoint correspondences are vital for surgical vision tasks like camera pose estimation, but their reliability is often compromised in surgical settings.

Purpose of the Study:

• To systematically evaluate state-of-the-art image matching techniques in laparoscopic surgery.
• To optimize image matching pipelines for resource-constrained surgical environments.
• To assess the impact of robust estimators on pose estimation accuracy.

Main Methods:

• Developed an optimized evaluation pipeline for image matching in surgical settings.
• Incorporated robust estimators for enhanced correspondence filtering and outlier resilience.
• Refined Mean Reprojection error calculation by thresholding nearest ground truth projections.
• Evaluated five robust estimators, identifying FM_8PTS as highly resilient to outliers.

Main Results:

• RoMa demonstrated superior performance in balancing pose estimation accuracy, reprojection performance, and computational efficiency.
• FM_8PTS showed the best resilience against outliers among the tested robust estimators.
• Optimized pipeline components, especially robust estimators, significantly influenced overall system performance.

Conclusions:

• RoMa is identified as the leading model for real-time surgical applications requiring accurate image matching.
• This work provides the first systematic benchmark for image matching in surgical domains.
• The findings pave the way for more reliable and efficient image-guided tools in minimally invasive surgery.