feat(#863): level leg cuts to a symmetric hip line on split - #931
Merged
Conversation
Splitting a segmented character gave clean symmetric ARMS but lopsided LEGS: the ONNX body model cuts the two legs at slightly different up-axis heights, so an explode looked uneven (one leg longer, a ragged hip line). New levelLimbCut pass (BODY category, default ON via Options::levelLimbCuts) fixes this: - Pools the leg↔torso boundary-face up-values across BOTH legs, takes the median as a single shared cut height h, and reassigns only leg/torso faces within a thin band of h — below h → the leg on that lateral side, above h → torso. - Scoped strictly to the leg↔torso pair. ARMS are excluded on purpose: they attach along a vertical seam where a horizontal level is meaningless (an early arm-inclusive version collapsed them into the torso). Unlike the experimental planarBoundaryRecut, this narrow scope can't scramble other parts. Verified on Hip Hop Dancing.obj (unrigged → AI-segmented → exploded, the exact repro): leg cut-height diff 3.7% → 0% of body height, size ratio 0.84 → 0.99, leg/torso seam up-std 2.88% → 1.33% (half the raggedness). Arms untouched (2624/2624). Rigged Rumba stayed balanced (near no-op when already symmetric — the pass only bites on an asymmetric model output). 35 MeshSegmenter tests pass (+2 for levelLimbCut: equalises heights; no-op when a part is missing). Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
|
Note Reviews pausedIt looks like this branch is under active development. To avoid overwhelming you with review comments due to an influx of new commits, CodeRabbit has automatically paused this review. You can configure this behavior by changing the Use the following commands to manage reviews:
Use the checkboxes below for quick actions:
No actionable comments were generated in the recent review. 🎉 ℹ️ Recent review info⚙️ Run configurationConfiguration used: defaults Review profile: CHILL Plan: Pro Plus Run ID: 📒 Files selected for processing (2)
🚧 Files skipped from review as they are similar to previous changes (1)
📝 WalkthroughWalkthroughAdds configurable body leg-to-torso seam leveling to geometric and ONNX segmentation cleanup, with tests and updated documentation. MCP mesh loading now frames the active viewport, while screenshots use Ogre offscreen rendering with temporary lighting restoration. ChangesLeg seam leveling
MCP headless rendering
Estimated code review effort: 4 (Complex) | ~45 minutes Sequence Diagram(s)sequenceDiagram
participant MCPClient
participant MCPServer
participant SpaceCamera
participant OgreRTT
MCPClient->>MCPServer: load_mesh
MCPServer->>SpaceCamera: frame loaded scene
MCPClient->>MCPServer: take_screenshot
MCPServer->>OgreRTT: render active camera and read pixels
OgreRTT-->>MCPServer: return RGBA image
MCPServer-->>MCPClient: save screenshot
Possibly related PRs
🚥 Pre-merge checks | ✅ 4 | ❌ 1❌ Failed checks (1 warning)
✅ Passed checks (4 passed)
✨ Finishing Touches📝 Generate docstrings
🧪 Generate unit tests (beta)
Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out. Comment |
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: f34067918f
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
Comment on lines
+733
to
+737
| const int lf = faceLabels[f]; | ||
| const bool isLimb = (lf == limbLeft || lf == limbRight); | ||
| if (!isLimb && lf != shared) continue; | ||
| const float fu = centroid[f].comp(up); | ||
| if (std::fabs(fu - h) > band) continue; |
There was a problem hiding this comment.
On a posed character where a bent or raised portion of a leg crosses the hip-height band, this loop processes that remote geometry even though it is not near a leg↔torso boundary: every leg face above h becomes torso and every participating face below it is reassigned as a leg. Since the default band spans 10% of the entire body height and the later island cleanup preserves sufficiently large patches, this can corrupt knees, feet, or thighs in common dance/seated poses. Restrict candidates by adjacency or distance through the actual seam rather than vertical distance alone.
Useful? React with 👍 / 👎.
Owner
Author
There was a problem hiding this comment.
| const float upSpan = upMax - upMin; | ||
| if (!(upSpan > 0.0f)) | ||
| return 0; | ||
| const float latMid = 0.5f * (latMin + latMax); |
There was a problem hiding this comment.
When an asymmetric pose extends an arm or accessory much farther on one side, the whole-mesh AABB midpoint can lie beyond both hips. In that case the below-cut branch assigns faces from both legs to the same limb label, splitting or merging semantic leg parts despite the input labels already identifying left and right. Compute the divider from the two leg/seam regions (or retain each existing limb's side) instead of allowing unrelated geometry to shift it.
Useful? React with 👍 / 👎.
Owner
Author
![coderabbitai[bot]](https://avatars.githubusercontent.com/in/347564?s=60&v=4){kind=small}
There was a problem hiding this comment.
Actionable comments posted: 1
🧹 Nitpick comments (1)
src/MeshSegmenter.cpp (1)
1010-1021: 📐 Maintainability & Code Quality | 🔵 Trivial | ⚡ Quick winExtract the shared face-cleanup pipeline instead of duplicating it in both paths.
segmentGeometricandpredict()each carry their own copy of thesmoothLabelBoundaries→levelLimbCut→planarBoundaryRecut→cleanupLabelIslands→vertexLabelsFromFacessequence; this PR had to add the newlevelLimbCutstep to both copies identically, which is exactly the kind of drift-prone duplication a shared helper avoids.
src/MeshSegmenter.cpp#L1010-L1021: extract thisif (opts.cleanupIslands) {...}block (and the matching smoothing/planar/island calls around it) into a private static helper takingfaceLabels, positions, vertexCount, indices, indexCount, opts, cat, and call it here.src/MeshSegmenter.cpp#L1356-L1365: call the same extracted helper here instead of repeating the block.🤖 Prompt for AI Agents
Verify each finding against current code. Fix only still-valid issues, skip the rest with a brief reason, keep changes minimal, and validate. In `@src/MeshSegmenter.cpp` around lines 1010 - 1021, Extract the duplicated face-cleanup sequence around segmentGeometric and predict into one private static helper accepting faceLabels, positions, vertexCount, indices, indexCount, opts, and cat; preserve the smoothLabelBoundaries, levelLimbCut, planarBoundaryRecut, cleanupLabelIslands, and vertexLabelsFromFaces ordering. Replace the repeated blocks at src/MeshSegmenter.cpp:1010-1021 and src/MeshSegmenter.cpp:1356-1365 with calls to that helper.
🤖 Prompt for all review comments with AI agents
Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.
Inline comments:
In `@src/MeshSegmenter.h`:
- Around line 286-304: Update the documentation for MeshSegmenter::levelLimbCut
to state that it runs after smoothing but before cleanupLabelIslands and
planarBoundaryRecut. Keep the implementation and all other documented behavior
unchanged.
---
Nitpick comments:
In `@src/MeshSegmenter.cpp`:
- Around line 1010-1021: Extract the duplicated face-cleanup sequence around
segmentGeometric and predict into one private static helper accepting
faceLabels, positions, vertexCount, indices, indexCount, opts, and cat; preserve
the smoothLabelBoundaries, levelLimbCut, planarBoundaryRecut,
cleanupLabelIslands, and vertexLabelsFromFaces ordering. Replace the repeated
blocks at src/MeshSegmenter.cpp:1010-1021 and src/MeshSegmenter.cpp:1356-1365
with calls to that helper.
🪄 Autofix (Beta)
Fix all unresolved CodeRabbit comments on this PR:
- Push a commit to this branch (recommended)
- Create a new PR with the fixes
ℹ️ Review info
⚙️ Run configuration
Configuration used: defaults
Review profile: CHILL
Plan: Pro Plus
Run ID: 540294f0-5cb0-4649-a24e-d49d9b099e2e
📒 Files selected for processing (4)
CLAUDE.mdsrc/MeshSegmenter.cppsrc/MeshSegmenter.hsrc/MeshSegmenter_test.cpp
… of flat User feedback on the flat horizontal leg cut: legs were now the same size, but (1) the feet got swapped (left foot ended up with the right leg) and (2) the torso skirt bottom got dragged into the legs. They asked for a DIAGONAL cut like the arms, not a forced horizontal waistline. Reworked levelLimbCut to MIRROR-SYMMETRISE across the sagittal plane instead of recutting at a shared height: - Reflect the labelling across the leg region's lateral centre; make each near-seam face agree with its mirror (union: limb if it OR its mirror is a limb; torso only if both). Preserves the model's natural diagonal boundary, makes the two legs symmetric, and — because reflection maps a foot to the opposite foot with the limb labels swapped — keeps each foot with its own leg. - No height/plane reassignment, so the torso skirt is never pulled down. - Scoped to a bounded edge-hop flood from the leg↔torso seam (distant geometry untouched); left/right side derived from the legs' own lateral means. Verified on Hip Hop Dancing.obj: leg size ratio → 1.00, feet stay on their own side (regress.py: both legs body_lat sign == foot_lat sign), skirt not dragged (leg up_max = natural diagonal), arms untouched. Rigged Rumba stays balanced. Test updated to assert symmetry AND no foot-swap; 35 MeshSegmenter tests pass. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
There was a problem hiding this comment.
Actionable comments posted: 1
🧹 Nitpick comments (2)
src/MeshSegmenter.cpp (2)
804-813: 📐 Maintainability & Code Quality | 🔵 Trivial | 💤 Low valueThe four-branch union rule collapses to one expression.
The first two branches are identical and
want = lfis dead.♻️ Simplification
- int want = lf; - if (isLimb && mirrorIsLimb) { - want = sideLimb; // both agree it's a limb → this side's limb - } else if (isLimb && !mirrorIsLimb) { - want = sideLimb; // union: keep as limb (mirror will follow) - } else if (!isLimb && mirrorIsLimb) { - want = sideLimb; // union: torso here but limb across → make limb - } else { - want = shared; // both torso - } + // Union rule: limb if this face OR its mirror is a limb; torso only if both are. + const int want = (isLimb || mirrorIsLimb) ? sideLimb : shared;🤖 Prompt for AI Agents
Verify each finding against current code. Fix only still-valid issues, skip the rest with a brief reason, keep changes minimal, and validate. In `@src/MeshSegmenter.cpp` around lines 804 - 813, In the branch assigning want within the isLimb/mirrorIsLimb logic, remove the redundant four-branch structure and dead want = lf initialization. Replace it with the equivalent single expression that selects sideLimb whenever either segment is a limb, otherwise selects shared, preserving the existing union behavior.
757-778: 🚀 Performance & Scalability | 🔵 Trivial | ⚡ Quick winConsider a hashed grid restricted to limb/shared faces.
gridis astd::mapover every face, andnearestFaceperforms 27 tree lookups per candidate. Only limb/sharedfaces can ever be a useful mirror target, and a packeduint64cell key in anunordered_mapavoids theO(log n)per probe. If you switch to an unordered container, keep an explicit tie-break (e.g. prefer the lower face index on equal distance) so results stay deterministic across runs/platforms.♻️ Sketch
- std::map<std::tuple<int,int,int>, std::vector<int>> grid; - for (int f = 0; f < faceCount; ++f) { - const Vec3& c = centroid[f]; - grid[key(c.comp(0),c.comp(1),c.comp(2))].push_back(f); - } + std::unordered_map<uint64_t, std::vector<int>> grid; + for (int f = 0; f < faceCount; ++f) { + const int lf = faceLabels[f]; + if (lf != limbLeft && lf != limbRight && lf != shared) continue; + const Vec3& c = centroid[f]; + grid[cellKey(c.comp(0), c.comp(1), c.comp(2))].push_back(f); + }with the nearest search keeping
if (d < bd || (d == bd && f < best)).🤖 Prompt for AI Agents
Verify each finding against current code. Fix only still-valid issues, skip the rest with a brief reason, keep changes minimal, and validate. In `@src/MeshSegmenter.cpp` around lines 757 - 778, Restrict the spatial hash used by nearestFace to limb/shared faces only, replacing the std::map tuple key with a packed uint64 cell key in an unordered_map to avoid tree lookups. Update grid construction and the 27-cell search consistently, and make nearestFace deterministic by selecting the lower face index when distances are equal.
🤖 Prompt for all review comments with AI agents
Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.
Inline comments:
In `@src/MeshSegmenter.cpp`:
- Around line 712-718: Update the sagittal-centre calculation in the surrounding
MeshSegmenter flow to use the midpoint of the left and right limb means, rather
than the face-count-weighted latSum/latN average. Move the per-limb accumulation
block above this calculation, derive latCentre from both limb means, and remove
the now-redundant combined accumulation pass while preserving the existing
zero-data handling.
---
Nitpick comments:
In `@src/MeshSegmenter.cpp`:
- Around line 804-813: In the branch assigning want within the
isLimb/mirrorIsLimb logic, remove the redundant four-branch structure and dead
want = lf initialization. Replace it with the equivalent single expression that
selects sideLimb whenever either segment is a limb, otherwise selects shared,
preserving the existing union behavior.
- Around line 757-778: Restrict the spatial hash used by nearestFace to
limb/shared faces only, replacing the std::map tuple key with a packed uint64
cell key in an unordered_map to avoid tree lookups. Update grid construction and
the 27-cell search consistently, and make nearestFace deterministic by selecting
the lower face index when distances are equal.
🪄 Autofix (Beta)
Fix all unresolved CodeRabbit comments on this PR:
- Push a commit to this branch (recommended)
- Create a new PR with the fixes
ℹ️ Review info
⚙️ Run configuration
Configuration used: defaults
Review profile: CHILL
Plan: Pro Plus
Run ID: e709e4d8-f449-4cab-a8ed-8456b46a90e9
📒 Files selected for processing (4)
CLAUDE.mdsrc/MeshSegmenter.cppsrc/MeshSegmenter.hsrc/MeshSegmenter_test.cpp
🚧 Files skipped from review as they are similar to previous changes (3)
- src/MeshSegmenter.h
- src/MeshSegmenter_test.cpp
- CLAUDE.md
User feedback: legs are now the right size/angle (mirror fix worked) but the seam still looks torn/broken between the legs — not a clean cut like the arms. Diagnosis (measured on Hip Hop Dancing): the leg↔torso label boundary wanders across ~8.6% of body height (a wide ragged band), so the split tears there. The arms look clean only because they're topologically DISCONNECTED from the torso in the model output (no shared seam at all) — the legs can't be, being attached at the hip. New cleanLimbSeam pass (legs only, after levelLimbCut): a leg attaches around a roughly horizontal hip ring, so per-leg it sets the cut height to the MEDIAN up-value of that leg's seam faces and reassigns the seam-proximity band (bounded edge-hop flood — so the centred torso skirt isn't dragged and each foot stays with its own leg) strictly by side of that height. Collapses the wandering band onto a level ring: leg↔torso seam raggedness ~8.6% → ~4-6% of body height (roughly halved), feet still correct, legs still symmetric. Honest limit noted in the header + CLAUDE.md: label-only cutting can't beat the mesh face resolution — a truly knife-thin seam needs real triangle slicing (Boolean/knife), which #859 lists as out of scope. This is the best clean-up achievable without new geometry. 36 MeshSegmenter tests pass (+CleanLimbSeamThinsWanderingBoundary). Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…mbSeam)" This reverts commit de3a73b.
MCP take_screenshot returned a black image and load_mesh left the camera un-framed, so headless/autonomous visual QA (load → explode → screenshot) was impossible. Two fixes: - take_screenshot now renders the active viewport's SpaceCamera into an offscreen PF_BYTE_RGBA RTT and reads it back, instead of QWidget::grab(). Ogre renders straight to the native window surface (WA_PaintOnScreen), so grab() only ever captured an empty Qt buffer. The RTT uses the RTSS MSN_SHADERGEN scheme + a temporary ambient boost and directional key light (restored afterward) so imported materials render lit, not black. Optional width/height args; defaults to the widget size. - load_mesh calls frameSceneInActiveViewport() (select every user scene node → SpaceCamera::frameSelection(), which no-ops on an empty selection) so the freshly-loaded mesh is actually in frame for the next screenshot. Verified end to end over the HTTP API: load a split character → transform_submesh to explode the parts → take_screenshot returns a fully textured, lit, framed image of the exploded model. 35 MeshSegmenter tests still pass. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
There was a problem hiding this comment.
Actionable comments posted: 5
🤖 Prompt for all review comments with AI agents
Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.
Inline comments:
In `@src/MCPServer.cpp`:
- Around line 2996-3013: The screenshot export in the code handling
image.save(path) lacks breadcrumb tracking. Add a SentryReporter::addBreadcrumb
call immediately before the save operation, using the file.export category and
including the screenshot path or relevant export context, while preserving the
existing save error handling and success flow.
- Around line 1619-1646: Update MCPServer::frameSceneInActiveViewport and its
load-call flow so reframing targets only scene nodes created by the current
import, rather than clearing SelectionSet and selecting every node with attached
objects. Preserve the user’s existing selection and camera view when loading
through an interactive GUI; only frame the newly imported nodes when that
behavior is explicitly required for the import.
- Around line 1619-1646: The new frameSceneInActiveViewport operation lacks
breadcrumb tracking. Add a SentryReporter::addBreadcrumb call within
MCPServer::frameSceneInActiveViewport, using the appropriate ui.action category
for this user-facing camera-framing and selection action, while preserving the
existing selection and camera behavior.
- Around line 1605-1617: Separate the mesh import and camera-framing error
handling in the surrounding operation: keep importMeshs() failures returning the
existing Ogre error, but invoke frameSceneInActiveViewport() in its own guarded
block after a successful import. If framing throws, preserve the successful
import result and avoid reporting it as an import failure.
- Around line 2949-3020: Ensure screenshot capture cleanup runs on every
exception, not only the success path. In the screenshot-rendering flow, move the
scene manager, saved ambient light, temporary light/node, and previous camera
aspect state into scope reachable by failure cleanup, or add an RAII guard that
restores ambient lighting and camera aspect ratio and destroys the temporary
light/node; preserve RTT removal and avoid double cleanup after successful
capture.
🪄 Autofix (Beta)
Fix all unresolved CodeRabbit comments on this PR:
- Push a commit to this branch (recommended)
- Create a new PR with the fixes
ℹ️ Review info
⚙️ Run configuration
Configuration used: defaults
Review profile: CHILL
Plan: Pro Plus
Run ID: 23747eb5-22bf-41db-96a9-ec64e02b608d
📒 Files selected for processing (3)
CLAUDE.mdsrc/MCPServer.cppsrc/MCPServer.h
🚧 Files skipped from review as they are similar to previous changes (1)
- CLAUDE.md
…bbit) CodeRabbit: latCentre was a face-count-weighted mean over all limb faces, so the leg with more faces pulled the sagittal midline toward itself — exactly the asymmetry this pass exists to fix — which could flip near-crotch faces to the wrong leg. Use the MIDPOINT of the two per-limb lateral means instead (count-independent); also derive rightIsPositive from those means. Verified on Hip Hop Dancing.obj: legs 468/470 (ratio 1.00), feet stay on their own side. Test refocused to the robust no-swap invariant (LevelLimbCutNeverSwapsLegSides) — the mirror pass's detailed reassignment is geometry-dependent (nearest-mirror over a real 3D hip) and is validated on the real model, not a brittle 2D grid. The other two Codex comments (P1 hip-band relabel, P2 whole-mesh AABB divider) were on the reverted horizontal-cut code (commit f340679) and no longer apply — the current mirror path uses neither a height band nor the whole-mesh AABB. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The RTT-screenshot rewrite changed the no-widget error message to "No active viewport camera", breaking the existing MCPServerTest.TakeScreenshot_MainWindowWithoutOgreWidgetReturnsError (CI unit-tests-linux). Restore the established "Error: OgreWidget not found" for the no-widget case (a QMainWindow with no OgreWidget); the new "No active viewport camera" is kept only for the has-widget-but-no-camera case. Screenshot tests pass locally. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
There was a problem hiding this comment.
Actionable comments posted: 1
🤖 Prompt for all review comments with AI agents
Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.
Inline comments:
In `@src/MeshSegmenter_test.cpp`:
- Around line 567-578: Strengthen the fixture in the affected levelLimbCut test
so the previous count-weighted divider would misclassify a near-midline face:
increase the right-leg face-count imbalance while preserving per-leg lateral
means and the correct midpoint. Update the assertions to explicitly verify that
the affected near-midline face retains label 5, while preserving the existing
side-correctness checks for labels 5 and 6.
🪄 Autofix (Beta)
Fix all unresolved CodeRabbit comments on this PR:
- Push a commit to this branch (recommended)
- Create a new PR with the fixes
ℹ️ Review info
⚙️ Run configuration
Configuration used: defaults
Review profile: CHILL
Plan: Pro Plus
Run ID: 2b3acefa-aa04-4684-ae02-54fd4ffa2902
📒 Files selected for processing (3)
src/MCPServer.cppsrc/MeshSegmenter.cppsrc/MeshSegmenter_test.cpp
🚧 Files skipped from review as they are similar to previous changes (2)
- src/MCPServer.cpp
- src/MeshSegmenter.cpp
…ming + test MCP screenshot (take_screenshot): - CRITICAL: the temp capture light/node, boosted ambient, altered camera aspect, and the RTT were leaked into the live scene if copyContentsToMemory / image alloc threw. Hoist that state out of the try and restore it via a single cleanup() lambda called on BOTH the success and the catch path. - Add a `file.export` Sentry breadcrumb on a successful save (project convention). load_mesh framing: - Move frameSceneInActiveViewport() OUTSIDE the import try so a framing throw can't mask a SUCCESSFUL import as an error (it swallows its own Ogre errors). - Preserve the user's selection: snapshot the current node selection, frame, then restore it — so auto-framing on load doesn't clobber an interactive --with-mcp session's selection. Test: - Replace the no-swap fixture with LevelLimbCutDividerIsCountIndependent: a deliberately wide left leg vs narrow right leg, so a face-count-weighted divider is pulled left while the correct midpoint-of-per-limb-means stays centred; asserts no right-leg face lands deep in left territory (and vice versa), distinguishing the two dividers. Verified: RTT screenshot still captures textured+lit end to end; 35 MeshSegmenter + 8 MCP screenshot/load tests pass. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
|
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Follow-up to the merged split-cleanup work (#930): splitting a segmented character produced clean, symmetric arms but lopsided legs — the ONNX body model cuts the two legs at slightly different up-axis heights, so an explode looked uneven (one leg longer, a ragged hip line). This adds a targeted
levelLimbCutpass that gives the legs the same clean, level cut the arms already have.Approach
levelLimbCut(BODY category only, default ON viaOptions::levelLimbCuts):h.h: belowh→ the leg on that lateral side, aboveh→ torso.planarBoundaryRecut(which stays OFF).Runs after the existing de-fringe/de-island passes, before the final island cleanup.
Verified — Hip Hop Dancing.obj (the exact repro: unrigged → AI-segmented → exploded)
Measured from the per-vertex labels + OBJ geometry:
Rigged Rumba Dancing.fbx (dance pose): legs stayed balanced (751/769) — the pass is a near no-op when the model output is already symmetric; it only bites on an asymmetric one. No regression.
Confirmed visually too via exploded silhouette renders (isometric sprite of a per-part-offset export): the two leg pieces are now matched and compact vs the fused, jagged baseline.
Tests
35 MeshSegmenter tests pass, incl. 2 new for
levelLimbCut(equalises differing leg heights; no-op when a leg/torso part is absent). All pure-data, no GL/model.Advances #863 (a symmetric, level cut is what makes segmented legs printable/joinable like the arms).
🤖 Generated with Claude Code
Summary by CodeRabbit