Fix #117367: OBJ: fix vertex colors imported scrambled up #126065

Scurest · 2024-08-08T00:08:40+02:00

Scurest commented

2024-08-08 00:08:40 +02:00

Lift certain incorrect assumptions about the order of vertices in
an OBJ file when processing vertex colors, which could lead to
missing or randomly permuted colors.

Replaces the list-of-blocks representation, attuned to assumptions
that the verts in an object form a contiguous subrange, with a
flat array better suited for random-access.

Recall that an OBJ has a global list of vertices (v), each of which may or may not have a vertex color. Each object (o) references a subset of vertices in this global list. For each referenced vert, the importer creates one Blender vert in the imported mesh. The vertex index in the global OBJ list is the global index, and the index in the Blender mesh's list of verts is the local index.

Specifically, the incorrect assumption in the importer appears to be that an object always references a contiguous range of vertices in the global list (no gaps) between the first and last referenced verts, and therefore (1) whether it has vertex colors just depends on if that range lies in a range that has vertex colors, and (2) the local index for a vert is just the distance to the start of the range in the global list.

(Note that vertex positions are stored the same way as colors, but the code for them does not make this assumption, so it's okay.)

So this does the following:

Replaces the list-of-blocks representation with a flat vertex color array. -1 is used to mark verts that did not have color. As an optimization, the array can omit trailing elements if they are all missing colors (so no memory cost for files without vertex colors). This is less sparse, but simpler for random-access.
Does a full scan of all referenced vertices to determine if they all have colors. As opposed to the coarse-but-fast range check.
Does a second scan to actually fill out the colors via color_attr_data[local_index] = global_vertex_colors[global_index]. Reusing the global-to-local mapping from create_vertices. This is the part that actually fixes the misaligned color problem.

In terms of performance, there should be no cost to importing files without vertex colors. Others are expected to slow down.

Fixes all the files posted in #117367 in my tests.

Lift certain incorrect assumptions about the order of vertices in an OBJ file when processing vertex colors, which could lead to missing or randomly permuted colors. Replaces the list-of-blocks representation, attuned to assumptions that the verts in an object form a contiguous subrange, with a flat array better suited for random-access. ------- Recall that an OBJ has a global list of vertices (`v`), each of which may or may not have a vertex color. Each object (`o`) references a subset of vertices in this global list. For each referenced vert, the importer creates one Blender vert in the imported mesh. The vertex index in the global OBJ list is the _global index_, and the index in the Blender mesh's list of verts is the _local index_. Specifically, the incorrect assumption in the importer appears to be that an object always references a contiguous range of vertices in the global list (no gaps) between the first and last referenced verts, and therefore (1) whether it has vertex colors just depends on if that range lies in a range that has vertex colors, and (2) the local index for a vert is just the distance to the start of the range in the global list. (Note that vertex positions are stored the same way as colors, but the code for them does not make this assumption, so it's okay.) So this does the following: * **Replaces the list-of-blocks representation with a flat vertex color array**. -1 is used to mark verts that did not have color. As an optimization, the array can omit trailing elements if they are all missing colors (so no memory cost for files without vertex colors). This is less sparse, but simpler for random-access. * **Does a full scan of all referenced vertices to determine if they all have colors**. As opposed to the coarse-but-fast range check. * **Does a second scan to actually fill out the colors via `color_attr_data[local_index] = global_vertex_colors[global_index]`**. Reusing the global-to-local mapping from `create_vertices`. This is the part that actually fixes the misaligned color problem. In terms of performance, there should be no cost to importing files without vertex colors. Others are expected to slow down. Fixes all the files posted in #117367 in my tests.

❤️ 1

Iliya Katushenock added this to the Pipeline, Assets & IO project 2024-08-08 00:49:05 +02:00

Aras Pranckevicius commented

2024-08-08 09:38:09 +02:00

@blender-bot build

Aras Pranckevicius reviewed 2024-08-08 09:39:52 +02:00

source/blender/io/wavefront_obj/importer/obj_import_mesh.cc Outdated

						
				@ -440,0 +423,4 @@

				  /* First pass to determine if we need to create a color attribute.

				   * Only when all vertices used specify a vertex color. */

				  for (int vi : mesh_geometry_.vertices_) {

				    const float3 color = global_vertices_.get_vertex_color(vi);

Aras Pranckevicius commented

2024-08-08 09:39:52 +02:00

Could this check be made a bit faster by having something like bool has_vertex_color(vi) that simply returns true or false? Right now this creates a whole "dummy" float3 only to check the validity.

And then seemingly whole get_vertex_color is not even actually needed elsewhere

Could this check be made a bit faster by having something like `bool has_vertex_color(vi)` that simply returns true or false? Right now this creates a whole "dummy" float3 only to check the validity. And then seemingly whole `get_vertex_color` is not even actually needed elsewhere

Scurest commented

2024-08-08 13:52:55 +02:00

How's that?

Aras Pranckevicius commented

2024-08-08 17:33:29 +02:00

Good! Does your latest push fix the obj tests by chance? They were all crashing on the buildbot this morning.

Scurest commented

2024-08-08 17:47:11 +02:00

If by crashing you mean the tests failed, no that is not fixed.

Aras Pranckevicius commented

2024-08-08 19:18:33 +02:00

Since you force-pushed it's hard to see the buildbot failures on the previous commit. But IIRC the output of the failing test simply said "segfault", which I assume means something within the test crashed. Which makes them fail, yes.

aras_p marked this conversation as resolved

Scurest force-pushed fix-117367 from 599ef78a25 to 8c5148042b

2024-08-08 13:51:16 +02:00

Compare

Scurest commented

2024-08-08 17:46:47 +02:00

So I have broken #MRGB colors. AFAICT they work like this

v ... # v1
v ... # v2
#MRGB ffff0000  # sets color for v1 (-2)
#MRGB ffff0000  # sets color for v2 (-1)
v ... # v3
v ... # v4
#MRGB ffff0000  # sets color for v4 (-1)

This is easier to handle with the list-of-blocks representation (perhaps that's why it was used?), since you don't know when you process each #MRGB line what vertex index it affects until you have seen the rest of the #MRGB lines up to the next v (or EOF).

My current idea is to buffer one Vector<float3> mrgb_block in GlobalVertices, accumulate into it on every #MRGB line, and flush it out the to the vertex_colors list at each v or EOF. How does that sound?

So I have broken #MRGB colors. AFAICT they work like this ``` v ... # v1 v ... # v2 #MRGB ffff0000 # sets color for v1 (-2) #MRGB ffff0000 # sets color for v2 (-1) v ... # v3 v ... # v4 #MRGB ffff0000 # sets color for v4 (-1) ``` This is easier to handle with the list-of-blocks representation (perhaps that's why it was used?), since you don't know when you process each #MRGB line what vertex index it affects until you have seen the rest of the #MRGB lines up to the next `v` (or EOF). My current idea is to buffer one `Vector<float3> mrgb_block` in `GlobalVertices`, accumulate into it on every #MRGB line, and flush it out the to the `vertex_colors` list at each `v` or EOF. How does that sound?

Aras Pranckevicius commented

2024-08-08 19:17:16 +02:00

This is easier to handle with the list-of-blocks representation (perhaps that's why it was used?)

Quite probably. I forget which order I implemented those in, but possibly I first tried to make MRGB work, and that is indeed very much "block based", and quite poorly defined (as in, it is not clear at all what should happen when multiple objects are present, etc.). And as you figured out, I did some poorly thought out decisions in handling various xyzrgb cases.

My current idea <...> How does that sound?

Sounds sensible to me.

> This is easier to handle with the list-of-blocks representation (perhaps that's why it was used?) Quite probably. I forget which order I implemented those in, but possibly I first tried to make MRGB work, and that is indeed very much "block based", and quite poorly defined (as in, it is not clear at all what *should happen* when multiple objects are present, etc.). And as you figured out, I did some poorly thought out decisions in handling various xyzrgb cases. > My current idea <...> How does that sound? Sounds sensible to me.

Scurest added 1 commit 2024-08-08 21:38:45 +02:00

Fix: OBJ: fix importing of #MRGB vertex colors

buildbot/vexp-code-patch-lint Build done.

Details

buildbot/vexp-code-patch-linux-x86_64 Build done.

Details

buildbot/vexp-code-patch-darwin-x86_64 Build done.

Details

buildbot/vexp-code-patch-darwin-arm64 Build done.

Details

buildbot/vexp-code-patch-windows-amd64 Build done.

Details

buildbot/vexp-code-patch-coordinator Build done.

Details

b3f618f0af

Fix non-trivial OBJs with #MRGB colors by buffering a block of
MRGB lines and flushing them to the main vertex_colors array.

Scurest commented

2024-08-09 02:08:05 +02:00

Okay, the mrgb_block is implemented. Tests are passing locally.

Okay, the `mrgb_block` is implemented. Tests are passing locally.

Aras Pranckevicius commented

2024-08-09 07:41:17 +02:00

@blender-bot build

Aras Pranckevicius approved these changes 2024-08-09 08:51:51 +02:00

Aras Pranckevicius merged commit 86e7668b11 into main

2024-08-09 08:54:08 +02:00

Aras Pranckevicius referenced this issue from a commit

2024-08-09 08:54:15 +02:00

Fix #117367: OBJ importer does not correctly import vertex colors when not all vertices have them

Aras Pranckevicius commented

2024-08-09 08:55:40 +02:00

Excellent, thanks! Merged to main for 4.3, also added an item for backporting into 4.2.2.

👍 1

Aras Pranckevicius referenced this issue from a commit

2024-08-09 13:50:19 +02:00

Tests: extend OBJ test coverage for #126065

Scurest deleted branch fix-117367

2024-08-09 17:08:34 +02:00

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 #117367: OBJ: fix vertex colors imported scrambled up #126065