Colmap utility functions and editing colmap functions for MP-SfM #3331
Conversation
| @@ -343,18 +343,19 @@ size_t ObservationManager::FilterObservationsWithNegativeDepth() { | |||
| size_t ObservationManager::FilterPoints3DWithSmallTriangulationAngle( | |||
| const double min_tri_angle, | |||
| const std::unordered_set<point3D_t>& point3D_ids) { | |||
| const std::unordered_set<point3D_t> to_remove = FindPoints3DWithSmallTriangulationAngle(min_tri_angle, point3D_ids); | |||
| const std::vector<point3D_t> ids_vec(point3D_ids.begin(), point3D_ids.end()); | |||
There was a problem hiding this comment.
@ahojnnes updated FindPoints3DWithSmallTriangulationAngle inputs/outputs following #3331 (comment). However, I did not convert the FilterPoints3DWithSmallTriangulationAngle input to std::unordered_set<point3D_t>.
As a result, I need to convert ids from set to vec. Changing the input to vec as well would also mean changing the inputs to this function in the COLMAP pipeline
| "options"_a, | ||
| "image_id"_a, | ||
| "reconstruction"_a); | ||
| &IncrementalMapper::ClearModifiedPoints3D) |
There was a problem hiding this comment.
@ahojnnes done following #3331 (comment)
Note that find_local_bundle is taken (uses the reconstruction object in self), hence the new name. I need to find a local bundle for an input reconstruction. I've made it static because I don't use IncrementalMapper in MP-SfM. This why I can do pycolmap.IncrementalMapper.find_rec_local_bundle(options, refimid, self.rec).
There was a problem hiding this comment.
Maybe find_local_bundle_from_reconstruction? Your current naming does not give intuitive meaning.
There was a problem hiding this comment.
OK, I guess I didn't look close enough in that this takes a custom 3D reconstruction and not the one managed by the mapper? Sorry for the confusion, but in this case a free-standing function could indeed be more suitable.
| @@ -139,6 +139,9 @@ class Reconstruction { | |||
| // Delete all 2D points of all images and all 3D points. | |||
| void DeleteAllPoints2DAndPoints3D(); | |||
|
|
|||
| // Return the 3D coordinates of the given point IDs. | |||
| Eigen::MatrixXd Point3DCoords(const std::vector<point3D_t>& point3D_ids) const; | |||
There was a problem hiding this comment.
reconstruction.Point3DCoords and image.Point2DCoords should ideally have identical interfaces.
There was a problem hiding this comment.
@ahojnnes @B1ueber2y I agree, however, I don't think there is an elegant solution for this one. The biggest differences is that image.Point2DCoords returns all point2D coordinates, while reconstruction.Point3DCoords takes list of 3D points. Since COLMAP filters 3D points, not all ids between 1 and the maximum point3D id has a 3D point. Thus, if we omit the input ids, we either
- return as many coordinates as there are 3D points, such that they each correspond to a sorted list of ids, or
- we return as many coordinates as the max point3D id, with some filler value where there is a missing 3D point. This way, in python, we could access the point using a 1D array of point3D ids (like for
image.Point2DCoords)
| } | ||
|
|
||
| void BindIncrementalMapper(py::module& m) { | ||
| BindIncrementalMapperImpl(m); | ||
| BindIncrementalPipeline(m); | ||
|
|
||
| m.def("find_local_bundle_from_reconstruction", |
There was a problem hiding this comment.
@ahojnnes following your suggestion #3331 (comment), I have made this into a free-standing function. What do you think of this placement?
|
Seems like only #3331 (comment), #3331 (comment) and #3331 (comment) remain |
Splitting my last PR into 2 Part 1.
These functions are needed for MP-SfM. They are also generic extensions of COLMAP that could be used elsewhere.
See original PR #3269