FM-Net: Deep Learning Network for the Fundamental Matrix Estimation from Biplanar Radiographs

Area of Science:

• Computer Vision
• Medical Imaging Analysis

Background:

• Fundamental matrix estimation is crucial in computer vision but hindered by the difficulty of accurate correspondence matching in biplanar radiographs.
• Traditional methods demand high-precision correspondences, which are challenging to obtain from biplanar radiographs.

Purpose of the Study:

• To develop an end-to-end deep learning network for direct fundamental matrix estimation from biplanar radiographs.
• To address the lack of publicly available datasets for biplanar radiograph analysis by creating a new dataset.

Main Methods:

• An end-to-end network, FM-Net, was designed for feature extraction and regression prediction to estimate the fundamental matrix directly from biplanar radiographs.
• A novel dataset of biplanar radiographs was generated for training and evaluating the proposed network.
• Performance was measured using Mean Square Error, R-squared, Square Value of Extreme Constraint, and Absolute Value of Extreme Constraint.

Main Results:

• The proposed FM-Net achieved a Square Value of Extreme Constraint of 0.20 and an Absolute Value of Extreme Constraint of 0.43.
• The estimation accuracy of FM-Net demonstrated a significant improvement of over 53.53% compared to existing methods.
• Experimental results validated the network's capability in fundamental matrix estimation.

Conclusions:

• The developed network successfully estimates the fundamental matrix from biplanar radiographs.
• FM-Net surpasses the performance of classical algorithms and other deep learning-based approaches in this domain.