Fix #116183: Mirror sometimes chooses the wrong vertex #116202

Germano Cavalcante · 2023-12-14T19:08:20+01:00

2023-12-14 19:08:20 +01:00

The non-topology option of EDBM_verts_mirror_cache_begin_ex only uses
the vertices position to determine whether it is a mirror. However,
some vertices may have the same position. In these cases, another value
can be used to compare.

This commit proposes using the normal and connected edges of a vertex
to determine the best vertex in a group of vertices with the same
position.

Issue

Test file: bugReport.blend

The non-topology option of `EDBM_verts_mirror_cache_begin_ex` only uses the vertices position to determine whether it is a mirror. However, some vertices may have the same position. In these cases, another value can be used to compare. This commit proposes using the normal and connected edges of a vertex to determine the best vertex in a group of vertices with the same position. --- ## Issue <video src="/attachments/6206bf21-8f43-4e0e-9a0d-d806ab915879" controls></video> Test file: [bugReport.blend](attachments/0f74ce9c-28f1-4e6b-83ed-e8e57da6f405)

Germano Cavalcante added the

Module

Modeling

label 2023-12-14 19:08:20 +01:00

Germano Cavalcante added 1 commit 2023-12-14 19:08:27 +01:00

8d9f4e2b90 Fix #116183 : Mirror sometimes chooses the wrong vertex

The non-topology option of `EDBM_verts_mirror_cache_begin_ex` only uses
the vertices position to determine whether it is a mirror. However,
some vertices may have the same position. In these cases, another value
can be used to compare.

This commit proposes using the normal and connected edges of a vertex
to determine the best vertex in a group of vertices with the same
position.

Germano Cavalcante requested review from Campbell Barton 2023-12-14 19:08:48 +01:00

Campbell Barton commented

2023-12-15 08:59:04 +01:00

This is interesting but I think it could be approached differently since this is adding involved logic to support a fairly uncommon use case. Currently mirror just doesn't support exactly overlapping vertices, so #116183 isn't a bug. I checked the user manual and this could be made more explicit.

There are some things to consider:

Even if this is supported in edit-mode, it wont be supported in weight-paint mirroring, so if we support this it should eventually be supported there too.
There would be rare cases where there are many vertices with the range, while unlikely it means each lookup could potentially allocate a large array, which is then inspected using expensive operations.
I'd expect the common case to be slower on high poly meshes because BLI_kdtree_3d_range_search is more expensive and allocates an array (even for a single item).

Having said all this - the functionality would be nice to support.

Suggestions:

2nd pass for un-assigned vertices

When some vertices don't have mirrored vertices (and aren't axis-aligned),
perform a second pass.

Loop over the vertices, checking for missing indices.
When found, perform a nearest lookup for vertices at this vertices location, filtering to include only vertices that have a mirrored location (API support would need to be added).
If found, a more expensive tests can be performed to check if this vertex is a better candidate (tests similar to the ones in this PR can work).

This method has the advantage that there is practically no overhead unless there are vertices without matches (only detecting when there are vertices that have no match).
Avoiding performance regressions for users with high-poly, symmetrical meshes.

Add dimensions to the KD-tree

A way to avoid gathering vertices within a range, then performing a search on those vertices could be to add additional dimensions to the KD tree: e.g. the normal / combined-edge-location could be added and the current logic of performing a nearest lookup can be kept.

The extra dimensions would need to be scaled down so the location would be dominant.

The advantage with this is it will have predictable performance & can be more easily implemented for Mesh data (not just BMesh).

A possible disadvantage is unwanted bias from the extra dimensions that could result in issues choosing between very near vertices, although this issue exists with any method that favors other attributes over nearest location.

It will also be slower, even in situations where it's not needed.

For this reason it could be optional, although I'm also not that keen on adding more options.

This is interesting but I think it could be approached differently since this is adding involved logic to support a fairly uncommon use case. Currently mirror just doesn't support exactly overlapping vertices, so #116183 isn't a bug. I checked the user manual and this could be made more explicit. There are some things to consider: - Even if this is supported in edit-mode, it wont be supported in weight-paint mirroring, so if we support this it should eventually be supported there too. - There would be rare cases where there are many vertices with the range, while unlikely it means each lookup could potentially allocate a large array, which is then inspected using expensive operations. - I'd expect the common case to be slower on high poly meshes because BLI_kdtree_3d_range_search is more expensive and allocates an array (even for a single item). ---- Having said all this - the functionality would be nice to support. **Suggestions:** ### 2nd pass for un-assigned vertices When some vertices don't have mirrored vertices (and aren't axis-aligned), perform a second pass. - Loop over the vertices, checking for missing indices. - When found, perform a nearest lookup for vertices at this vertices location, filtering to include only vertices that have a mirrored location (API support would need to be added). - If found, a more expensive tests can be performed to check if this vertex is a better candidate (tests similar to the ones in this PR can work). This method has the advantage that there is practically no overhead *unless* there are vertices without matches *(only detecting when there are vertices that have no match).* Avoiding performance regressions for users with high-poly, symmetrical meshes. ---- ### Add dimensions to the KD-tree A way to avoid gathering vertices within a range, then performing a search on those vertices could be to add additional dimensions to the KD tree: e.g. the normal / combined-edge-location could be added and the current logic of performing a nearest lookup can be kept. The extra dimensions would need to be scaled down so the location would be dominant. The advantage with this is it will have predictable performance & can be more easily implemented for Mesh data (not just BMesh). A possible disadvantage is unwanted bias from the extra dimensions that could result in issues choosing between very near vertices, although this issue exists with any method that favors other attributes over nearest location. It will also be slower, even in situations where it's not needed. For this reason it could be optional, although I'm also not that keen on adding more options.

Merge conflict checking is in progress. Try again in few moments.

View command line instructions.

Checkout

From your project repository, check out a new branch and test the changes.

git fetch -u fix_116183:mano-wii-fix_116183

git checkout mano-wii-fix_116183

Sign in to join this conversation.

No reviewers

No Label

Download

What's New

Blender Studio

Manual

Developers Blog

Documentation

Benchmark

Blender Conference

Development Fund

One-time Donations

Fix #116183: Mirror sometimes chooses the wrong vertex #116202

Issue

2nd pass for un-assigned vertices

Add dimensions to the KD-tree

Checkout