feat: V2 stacking pipeline with OKLab MMCQ, depth ordering, and comprehensive optimization#3
Merged
Conversation
…hape extension V2 pipeline implementation (pipeline_v2.cpp) with: - OKLab MMCQ color quantization (src/quantize/) - Connected-component shape layers with depth ordering (src/stacking/) - Shape extension via morphological dilation into occluded regions - Per-layer Potrace tracing with path optimization (src/curve/path_optimize) - Z-order safe same-color shape merging - Diagnostic depth order validation logging Supporting changes: - PipelineMode enum and pipeline_mode config (config.h) - --pipeline CLI flag for raster_to_svg and evaluate_svg - Python bindings for V2 pipeline mode - is_hole flag on BezierContour for correct SVG winding - Eval: border_delta_e_mean metric, updated manifest - Research docs and cursor rules Made-with: Cursor
…acing Keep labels alive past shape tracing so ApplyCoverageGuard can identify uncovered regions. Patch shapes are rotated to the bottom of the z-order so they only fill through gaps without overpainting correct content. Made-with: Cursor
Depth ordering improvements: - Replace convex hull with morphological dilation for D(i,j) signal - Implement Tarjan SCC for efficient batch cycle breaking - Add adaptive delta threshold based on D(i,j) distribution (P25) - Improve background detection with border pixel frequency voting - Add significance filtering for small fragment pairs Eval rasterizer fix: - Replace heuristic hole detection in FillShapeWithHoles with correct SVG evenodd fill via XOR, matching browser rendering behavior Made-with: Cursor
- Increase K-Means refinement from 3 to 8 iterations with early convergence detection to produce more accurate centroids - Add post-MMCQ palette consolidation: greedily merge centroid pairs within OKLab distance < 0.025 to reduce over-segmentation - Add spatial label smoothing via 5x5 majority vote filter with color distance guard to eliminate salt-and-pepper noise at boundaries Made-with: Cursor
- Include Potrace-detected holes in VectorizedShape with is_hole flag, filtered by min_hole_area to avoid tiny artifacts. This fixes incorrect solid fill where shapes should have transparent regions (eval +1.3 pts). - Cap path optimization chain-merge at 3 segments to prevent a single cubic from representing too many original segments, which caused visible dents on smooth arcs. - Increase TryMergeSegments sample density from 8 to 12 per segment for more sensitive deviation detection. - Tighten merge_eps lower bound from 0.5 to 0.3 and coefficient from 0.6 to 0.5 for better curve fidelity. Eval: aggregate 68.0 -> 69.3 (+1.3). Notable gains: s_64c9a +11.5, c_5bf9c +5.5, tjls +3.6. One regression: c_193ce -3.8 (depth ordering errors exposed by holes). Made-with: Cursor
Phase 1: Remove dead code (CurveSegment, AngleBetween, FitBezierOnGraph, 3 boundary helpers, 7 assembly functions), invalid code (median_idx no-op), duplicate includes, and unnecessary includes. Phase 2: Extract shared utilities and eliminate duplication: - Unify Bezier eval functions and MakeLinearBezier across curve module - Extract LabDistSq/LabDist to cv_utils.h, replace 4 duplicate impls - Deduplicate OKLab SrgbToOklab via shared LinearRgbToOklab - Extract FindNearestCentroid in color_quantize.cpp - Deduplicate Potrace module (RingToBezier, RunPotraceTrace, SimplifyRing) - Extract MakeStrokeShape in thin_line.cpp - Extract ApplySubiteration in morphology.cpp ZhangSuenThinning Phase 3: Split long functions into focused sub-functions: - QuantizeColors -> RunMmcq + RefineCentroidsKMeans + ConsolidatePalette + SmoothLabels - ComputeDepthOrder -> BuildDepthGraph + AddContainmentEdges + TopologicalSortWithCycleBreaking - FindBackground -> CollectBorderInfo + SelectBackground - EstimateOptimalColors -> BuildColorSamples + ScoreCandidateK - ApplyCoverageGuard -> FindCoverageGaps + PatchMissingRegions - Extract pipeline shared logic (DeriveTraceParams, RescaleShapes, ClampShapesToBounds) - Extract VectorizeImpl from 3 Vectorize overloads Phase 4: Naming and style unification: - Rename ColorKey lambda to snake_case - Rename HistBin members to snake_case (sum_l, sum_a, etc.) - Unify V1 pipeline log prefixes to "V1: ..." format - Normalize same-module include paths (remove redundant subdir prefixes) Phase 5: Decouple VectorizedShape into src/detail/vectorized_shape.h, eliminating reverse dependencies from contour/curve/trace -> output. Made-with: Cursor
P0 - Hot path fixes: - Eliminate per-pixel heap allocation in SmoothLabels (freq vector reuse) - Cache OKLab values in ColorGrid::Build (single SrgbToOklab pass instead of 3) - Rewrite ExtractShapeLayers as single-pass O(WH) BFS (was O(L*WH)) - Cache per-label masks in V1 pipeline (eliminate repeated full-image compare) P1 - OpenMP parallelization (MSVC OpenMP 2.0 compatible): - Row-level parallel: pixel label assignment, SmoothLabels, ColorGrid histogram - Layer-level parallel: Potrace tracing in both V1 and V2 pipelines - Shape-level parallel: path optimization (OptimizeShapePaths) - Component-level parallel: coverage guard patch generation - V2 depth validation reduction(+:mismatch) P2 - Memory optimization: - Coverage guard: ROI-level masks instead of full-image allocation - ExtendShapeMasks: ROI-bounded dilation, pixel count tracker - ColorBox: replace priority_queue with make_heap/pop_heap (avoid deep copy) - RefineCentroidsKMeans: accumulator array reuse across iterations P3 - Algorithm & cache optimization: - SrgbToLinear: 256-entry runtime LUT for uint8 fast path - V2 depth validation: opt-in via enable_depth_validation config (default off) - ConsolidatePalette: Kruskal-style pre-sorted union-find (single pass) - EstimateOptimalColors: parallel candidate K evaluation All OpenMP pragmas use int loop variables and #ifdef _OPENMP guards for MSVC compatibility. Eval score unchanged at 69.3 (zero regression). Made-with: Cursor
- Export PipelineMode in Python __init__.py and __all__ - Add enable_depth_validation to PartialVectorizerConfig and MergeInto - Add --enable-depth-validation CLI flag to raster_to_svg and evaluate_svg - Fix V2 pipeline hardcoded stroke=false, now respects config - Add coverage guard step to AGENTS.md V2 flow diagram - Add V1/V2 applicability column to README config parameter table - Add coverage guard to README V2 feature list Made-with: Cursor
- Add cp314-* to cibuildwheel build targets - Add Python 3.14 to CI test matrix and pyproject.toml classifiers - Split Windows wheel builds into 3 parallel jobs (310-311, 312-313, 314) to reduce wall time (~3x speedup, matching neroued_3mf approach) - Upgrade cibuildwheel from v2.23 to v3.4.0 - Simplify macOS wheel to arm64-only (cross-compile x86_64 unreliable) - Use per-matrix cibw_build + id for clearer artifact naming - Update README supported Python versions to 3.10-3.14 Made-with: Cursor
M_PI is a POSIX extension not defined by MSVC unless _USE_MATH_DEFINES is set. Use C++20 std::numbers::pi instead. Made-with: Cursor
Add detailed MSVC-specific rules to prevent future issues: - Forbid M_PI and other POSIX math constants (use std::numbers) - Forbid constexpr with non-constexpr math functions on MSVC - Document OpenMP 2.0 limitations (int loop vars, no collapse/task) - Document POSIX function equivalents (_popen, localtime_s, etc.) - Forbid VLA and GCC/Clang extensions Made-with: Cursor
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
V2 层叠模型管线的完整实现，从颜色量化到 SVG 输出的端到端矢量化流程，以及全面的代码质量和性能优化。
V2 管线核心功能
质量评估
性能优化（4.18x 整体加速）
代码质量
Test plan
-O3验证--help所有参数正确显示PipelineMode可正常导入导出Made with Cursor