autonomousvision / monosdf

[NeurIPS'22] MonoSDF: Exploring Monocular Geometric Cues for Neural Implicit Surface Reconstruction
MIT License
572 stars 53 forks source link

"How to utilize COLMAP's trajectory for training." #102

Open liweishuo-bigai opened 2 weeks ago

liweishuo-bigai commented 2 weeks ago

When training on the Replica and ScanNet datasets, I found that using the datasets' built-in trajectories yielded high-quality meshes. However, replacing them with COLMAP output resulted in poorly reconstructed and incomplete meshes. I attempted to normalize the primary axis in the COLMAP poses, scaling the other two axes by the same factor, but the outcome remained chaotic. Additionally, I observed that adjusting the parameter “3” in the formula scale = 2. / (np.max(max_vertices - min_vertices) + 3.) influenced the results. Could you explain the purpose of this parameter? Do you have any suggestions for effectively incorporating COLMAP trajectories into the training process?