实验 21:A15 Protocol4 Ranker Retrospective

10-case 候选池路由回顾验证:拓扑候选有效,但几何保真仍需门控

实验定位 混合有效 / 回顾验证

A15 把 A14 的候选池思想扩展到已有 10-case protocol4 结果:候选池包含 baseline / selected-alpha / alpha025 / repair-hybrid。选择阶段显式不使用 GT、Chamfer 或 F-score,只使用生成状态、拓扑、面片数和尺度比例;GT 指标只在选择之后做事后评估。

关键限制:这批 10-case 数据没有每个候选的 shape-SLat preview,因此 A15 不是 A14 preview gate 的泛化证明,而是“候选池 + 无 GT proxy + topology candidate”能否扩展的回顾验证。

实验设计(Image-2 风格)

实验设计图:10-case protocol4 candidate-pool retrospective
实验设计图:10-case protocol4 candidate-pool retrospective
模块设计图:geometry candidate + topology candidate routing
模块设计图:geometry candidate + topology candidate routing

模块设计(Image-2 风格)

Hard Gate

  • candidate must have ok status
  • faces ratio <= 1.35
  • boundary ratio <= 1.40 unless tiny baseline boundary
  • nonmanifold ratio <= 1.35 unless tiny baseline nonmanifold
  • bbox ratio in [0.60, 1.55]
  • surface-area ratio in [0.45, 1.80]

Proxy Score

topology cleanup + compactness + watertight reward - bbox/surface-area drift; Chamfer/F-score/GT are used only for retrospective evaluation.

这一步的研究意义在于把“VGGT-guided 几何候选”和“repair topology 候选”放进同一个可路由候选池,而不是只做单一路线改进。

实验结果(表格)

方法成功平均 proxyChamferChamfer ΔF@5F@5 Δfacesfaces Δ水密率非流形边
baseline10/100.00000.08161-0.4336-950660-0.0%10348
selected_alpha9/100.03590.084170.001010.43140.00261048883-59060.0%10586
alpha0259/100.01700.083250.000090.42960.00081040042-147480.0%10125
repair_hybrid10/101.13120.081930.000320.43880.0052443226-507435100.0%0
ranker_selected10/101.05280.081670.000060.43960.0059406873-54378790.0%97

每个 case 的路由

caseranker 选择通过 hard gate 候选数Chamfer ΔF@5 Δfaces Δ水密
gso_000_input4repair_hybrid2-0.001600.0202-107621
gso_001_input4repair_hybrid4-0.001050.008230561
gso_002_input4repair_hybrid4-0.003770.0502-12275061
gso_003_input4repair_hybrid3-0.00005-0.0041-23305921
gso_004_input4repair_hybrid4-0.000410.0032-11853341
gso_005_input4repair_hybrid40.00119-0.007990521
gso_006_input4repair_hybrid40.001440.02043241441
gso_007_input4repair_hybrid40.00281-0.0188-2580281
gso_008_input4alpha02530.00209-0.0089-846300
gso_009_input4repair_hybrid4-0.00006-0.0029-6772741

选择来源统计

候选来源次数比例
alpha025110.0%
repair_hybrid990.0%

可视化结果

A15 Chamfer by method
A15 Chamfer by method
A15 F-score @5% by method
A15 F-score @5% by method
A15 predicted faces by method
A15 predicted faces by method
A15 non-manifold edges by method
A15 non-manifold edges by method
ranker Chamfer Δ
0.00006

正数表示比 baseline 略差;这是 topology 候选带来的几何保真代价。

ranker F@5 Δ
0.0059

F@5 均值提升,说明部分 surface 覆盖/精度改善。

ranker 水密率
90.0%

从 baseline 0% 提升到 90.0%,拓扑路线有效。

实验结论

下一步想法