Open nick0622 opened 3 months ago
Hi openMVS Team,
I hope this message finds you well. I wanted to follow up on my pull request (#1123) submitted on 3/25, which proposes an enhancement to the dense reconstruction process by allowing users to input critical points for better crack mapping.
Hi Nick! Thank you for the proposed enhancement, however it is not clear to me how you plan to use these points. Can you please elaborate on how you think they can improve densification? Usually accurate/sparse point matches are useful in SfM stage (before MVS) to better align the cameras. In MVS, if there is enough texture (as is the case for cracks), there is no matching ambiguity, and the generated point cloud should be accurate already.
Hi,
Upon reviewing the dense point cloud, it became evident that some crack points weren't adequately represented. This led me to consider incorporating additional information, such as manually marking correspondences between crack points across images. This approach aims to enhance the localization of these critical features in 3D space.
By providing these correspondences, the reconstruction algorithm can prioritize and refine the estimation of crack points' 3D coordinates. This is particularly valuable because not all crack points may be effectively matched during the initial SfM stage due to various factors like lighting variations or occlusions.
I believe integrating manual correspondences could offer a targeted improvement in capturing and accurately representing all critical crack points within the reconstructed model.
Looking forward to discussing this approach in more detail.
Best regards, Nick
hard to discuss without any examples of the issue, but again most probably the problem is inaccurate camera poses, which is in SfM, not here in MVS
Description
I am working on reconstructing a pier with visible cracks on its surface. The current dense reconstruction process does an excellent job of capturing the general structure but lacks the precision to accurately map the cracks visible in the input images. I propose a feature enhancement that allows users to input or mark specific points of interest (e.g., crack endpoints or critical points along the cracks) across different images that the reconstruction algorithm can then use to improve the fidelity of these features in the final model.
Suggested Implementation
Point Matching Input: Allow users to input pairs or groups of matched points across images, potentially through a simple GUI or a pre-defined format in a text file.
Feature Enhancement: Utilize these points as additional data during the point cloud generation phase to ensure these features are accurately represented.
Adjustment Parameters: Offer parameters to adjust the emphasis or weighting of these manually inputted points in the reconstruction process, allowing users to balance between general structure fidelity and detail accuracy.