实验 23:A17 View-consistency Silhouette Gate

把候选 mesh 渲染回输入非正交相机:silhouette 信号有效,但单独排序不足

实验定位 混合 / 方法可行但排序不足

A17 是 A16 之后的关键推进:不再只看 mesh 内部统计,而是把候选 mesh 渲染回输入的 4 个非正交相机,用输入前景 mask 与候选渲染 mask 的 silhouette IoU 作为无 GT 观测一致性信号。渲染分辨率为 64,选择阶段不使用 GT。

结果说明这个信号确实可计算,并且 silhouette_routed 的平均 IoU 最高;但单独优化 silhouette 仍不能保证 Chamfer/F-score 最优,也会选择非水密几何候选。因此下一步不是丢掉 silhouette,而是把它纳入 topology-hard + geometry-Pareto ranker。

实验设计(Image-2 风格)

实验设计图:candidate mesh 渲染回输入相机
实验设计图:candidate mesh 渲染回输入相机
模块设计图:view-consistency fidelity gate
模块设计图:view-consistency fidelity gate

模块设计(Image-2 风格)

输入信号

foreground masks estimated from the 4 input PNGs; candidate masks rendered by Blender from the same camera protocol

输入 mask 来自 alpha/非白像素自动估计,不使用人工标注和 GT mesh。

选择规则

mean silhouette IoU + watertight reward - area-ratio / topology / face-ratio penalties

max silhouette_score; GT metrics are used only for retrospective evaluation

实验结果(表格)

方法成功Silhouette IoUSilhouette scoreChamferChamfer ΔF@5F@5 Δfacesfaces Δ水密率
baseline10/100.46180.13070.08161-0.4336-950660-0.0%
ranker_selected10/100.46910.40890.081670.000060.43960.0059406873-54378790.0%
repair_hybrid10/100.46710.44460.081930.000320.43880.0052443226-507435100.0%
silhouette_routed10/100.47750.38340.081990.000390.43600.0024360611-59004980.0%
alpha0259/90.46570.12620.083250.000090.42960.00081040042-147480.0%
selected_alpha9/90.46080.12190.084170.001010.43140.00261048883-59060.0%

每个 case 的路由

case最终选择选中 IoUbaseline IoUIoU ΔChamfer ΔF@5 Δfaces Δ水密率
gso_000_input4repair_hybrid0.57830.43830.1400-0.001600.0202-10762100.0%
gso_001_input4repair_hybrid0.28290.2842-0.0013-0.001050.00823056100.0%
gso_002_input4repair_hybrid0.57830.55550.0228-0.003770.0502-1227506100.0%
gso_003_input4repair_hybrid0.74430.7444-0.0001-0.00005-0.0041-2330592100.0%
gso_004_input4repair_hybrid0.55840.5589-0.0006-0.000410.0032-1185334100.0%
gso_005_input4repair_hybrid0.34750.34750.00000.00119-0.00799052100.0%
gso_006_input4alpha0250.41440.4152-0.00080.00102-0.0101-1901620.0%
gso_007_input4repair_hybrid0.24600.24470.00130.00281-0.0188-258028100.0%
gso_008_input4selected_alpha0.24610.2475-0.00140.00577-0.0139-329400.0%
gso_009_input4repair_hybrid0.77910.7818-0.0027-0.00006-0.0029-677274100.0%

选择来源统计

候选来源次数比例
alpha025110.0%
repair_hybrid880.0%
selected_alpha110.0%

可视化结果

A17 silhouette IoU by method
A17 silhouette IoU by method
A17 Chamfer by method
A17 Chamfer by method
A17 watertight rate by method
A17 watertight rate by method
A17 predicted faces by method
A17 predicted faces by method
silhouette IoU
0.4775

高于 baseline 的 0.4618,说明观测一致性信号有效。

Chamfer Δ
0.00039

仍比 baseline 略差,也差于 A15 ranker 的 0.00006。

水密率
80.0%

因为选择了 2 个非水密几何候选,水密率低于 A15。

实验结论

下一步想法