Shaojie Zhang,Yinghui Wang,Bin Nan,Wei Li,Jinlong Yang,Tao Yan,Yukai Wang,Liangyi Huang,Mingfeng Wang,Ibragim R. Atadjanov
Abstract:To address the issue of increased triangulation uncertainty caused by selecting views with small camera baselines in Structure from Motion (SFM) view selection, this paper proposes a robust error-resistant view selection method. The method utilizes a triangulation-based computation to obtain an error-resistant model, which is then used to construct an error-resistant matrix. The sorting results of each row in the error-resistant matrix determine the candidate view set for each view. By traversing the candidate view sets of all views and completing the missing views based on the error-resistant matrix, the integrity of 3D reconstruction is ensured. Experimental comparisons between this method and the exhaustive method with the highest accuracy in the COLMAP program are conducted in terms of average reprojection error and absolute trajectory error in the reconstruction results. The proposed method demonstrates an average reduction of 29.40% in reprojection error accuracy and 5.07% in absolute trajectory error on the TUM dataset and DTU dataset.
What problem does this paper attempt to address?
The problem that this paper attempts to solve is in Structure from Motion (SFM) three - dimensional reconstruction, the problem of increased triangulation error due to the selection of views with a small baseline distance. Specifically, in the traditional SFM method, when selecting views, views with a smaller camera baseline are often selected based on content similarity or feature matching results, which will lead to an increase in the range of triangulation error, thereby reducing the accuracy of three - dimensional reconstruction. To solve this problem, the paper proposes an error - resistant view selection method based on the camera baseline. By calculating the triangulation error under different baselines, an error - resistant matrix is constructed, and this matrix is used to select a set of candidate views to improve the accuracy and efficiency of three - dimensional reconstruction.
### Main contributions of the paper:
1. **Propose an error - resistant view selection method based on the camera baseline**: This method can solve the problem of increased triangulation uncertainty caused by the selection of small - baseline views.
2. **Design a recursive view omission and supplementation strategy**: Ensure that all views can participate in three - dimensional reconstruction, thereby improving the accuracy of reconstruction.
3. **Integrate this method with the COLMAP application**: Experiments were carried out on the DTU and TUM datasets, and it was verified that this method has an average improvement of 29.40% in reprojection error compared to the exhaustive method of COLMAP.
### Method overview:
1. **Calculation of triangulation error resistance**:
- Quantify the error - resistant ability of the baseline by calculating the triangulation error caused by a 1 - pixel matching error under different baselines.
- Establish an error - resistant matrix, where each element represents the baseline error - resistant value between a pair of views.
2. **View selection**:
- Sort each row of the error - resistant matrix and select the views corresponding to the first five smallest error - resistant values as the set of candidate views.
- Recursively check and supplement views to ensure that all views participate in the reconstruction process.
### Experimental results:
- On the TUM and DTU datasets, compared to the exhaustive method of COLMAP, this method has an average reduction of 29.40% in reprojection error and a 5.07% reduction in absolute trajectory error.
### Key formulas:
- **Reprojection error**:
\[
E=\frac{1}{cm}\sum_{i = 1}^{m}\sum_{j = 1}^{n}\|x_{ij}-\pi(P_iX_j)\|^2_2
\]
where \(m\) is the number of views, \(n\) is the number of 3D points, \(c\) is the number of 2D points, \(p\) is the projection matrix, \(x_j\) is a 3D point, \(x_{ij}\) is the 2D coordinate of the 3D point in view \(i\), and \(\pi((x,y,z)^T)=(x/z,y/z)^T\) is the projection function.
- **Absolute trajectory error**:
\[
S'=\arg\min\sum_{i = 0}^{N - 1}\|P_i - sR(P'_i - t)\|^2
\]
where \(P_i\) is the true coordinate of camera position \(i\), \(P'_i\) is the estimated coordinate of camera position \(i\), \(s\) is the scaling parameter in the similarity transformation matrix, \(R\) is the rotation matrix, and \(t\) is the translation matrix.
Through these methods and experiments, the paper effectively solves the triangulation error problem caused by the selection of small - baseline views and improves the accuracy and efficiency of three - dimensional reconstruction.