The default parameters are optimal for most datasets. Also, many parameters are exposed for research & development purposes and are not useful for users. A subset of them can be useful for advanced users to improve the quality on specific datasets.
The first thing is to verify the number of reconstructed cameras from your input images. If a significant number are not reconstructed, you should focus on the options of the sparse reconstruction.
HighIf your dataset is not big (<300 images), you can use
Highpreset. It will take more time for the
StuctureFromMotionnode but it may help to recover more cameras. If you have really few images (like <50 images), you can also try
Ultrawhich may improve or decrease the quality depending on the image content.
Guided MatchingThis option enables a second stage in the matching procedure. After matching descriptor (with a global distance ratio test) and first geometric filtering, we retrieve a geometric transformation. The guided-matching use this geometric information to perform the descriptors matching a second time but with a new constraint to limit the search. This geometry-aware approach prevents early rejection and improves the number of matches in particular with repetitive structures. If you really struggle to find matches it could be beneficial to use
BRUTE_FORE_L2matching, but this is not good in most cases as it is very inefficient.
StructureFromMotionnodes It may improve especially on some surfaces (like skin for instance). It is also more affine invariant than SIFT and can help to recover connections when you have not enough images in the input.
To improve the robustness of the initial image pair selection/initial reconstruction, you can use a SfM with
minInputTrackLengthset to 3 or 4 to keep only the most robust matches (and improve the ratio inliers/outliers). Then, you can chain another SfM with the standard parameters, so the second one will try again to localize the cameras not found by the first one but with different parameters. This is useful if you have only a few cameras reconstructed within a large dataset.
DepthMapYou can adjust the
Downscaleparameter to drive precision/computation time. If the resolution of your images is not too high, you can set it to 1 to increase precision, but be careful, the calculation will be ~4x longer. On the contrary, setting it to a higher value will decrease precision but boost computation.Reduce the number of neighbour cameras (
SGM: Nb Neighbour Cameras,
Refine: Nb Neighbour Cameras) will directly reduce the computation time linearly, so if you change from 10 to 5 you will get a 2x speedup. A minimum value of 3 is necessary, 4 already gives decent results in many cases if the density of your acquisition process regular enough. The default value is necessary in a large scale environment where it is difficult to have 4 images that cover the same area.
DepthMapFilterIf you input images are not dense enough or too blurry and you have too many holes in your output. It may be useful to relax the
Min Consistent Camerasand
Min Consistent Cameras Bad Similarityto 2 and 3 respectively.
MeshingIf you have less than 16G of RAM, you will need to reduce the
Max Pointsto fit your RAM limits. You may also augment it, to recover a more dense/precise mesh.
Filter Large Triangles Factorcan be adjusted to avoid holes or on the other side to limit the number of large triangles.
Keep Only The Largest Mesh: Disable this option if you want to retrieve unconnected fragments that may be useful.
TexturingYou can change the
Texture Downscaleto 1 to improve the texture resolution.
You can choose to use one or multiple describer types. If you use multiple types, they will be combined together to help get results in challenging conditions. The values should always be the same between FeatureExtraction, FeatureMatching and StructureFromMotion. The only case, you will end up with different values is for testing and comparing results: in that case you will enable all options you want to test on the FeatureExtraction and then use a subset of them in Matching and SfM.