实验 67:A61 Soft MV Score Audit

后验 soft MV score 无法替代 A53,推动研究转向 generation-time 多视角约束

实验定位 后验 soft score 负结果 / 论文论证正向

A61 是 A60 后的上限审计:既然 hard veto 会错杀好编辑,就把 MV consistency 改成连续 soft score、worst-view 约束和 conservative defer,测试后验 mesh selector 是否仍能替代 A53。

结论很明确:不能。最好的 no-GT soft MV selector 是 a61_min_mean_mvd,mean ΔChamfer=0.0000166、ΔF@5=0.001037;A53 则为 ΔChamfer=-0.0004171、ΔF@5=0.002684。soft MV 比 A53 的 Chamfer 差 0.0004337,F@5 少 -0.001647。

实验设计(Image-2 风格)

实验设计图:A61 soft MV score audit
实验设计图:A61 soft MV score audit
模块设计图:post-hoc score vs generation-time budget
模块设计图:post-hoc score vs generation-time budget

模块设计(Image-2 风格)

Soft Scores

  • a61_min_mean_mvd:选择平均 MV depth error 最低的候选。
  • a61_min_worst_view:选择最坏输入视角 error 最低的候选。
  • a61_soft_pareto:mean、worst-view、std 和步数惩罚的连续组合。
  • a61_budgeted_soft_pareto / a61_conservative_defer:只接受不破坏 carrier MV budget 的候选,否则回退 carrier。

为什么这是关键审计

A61 没有引入 GT 参与选择,只把 GT 用作回顾评测。它检验的是后验选择器的上限:如果 soft MV score 仍无法超过 A53,下一步就不应继续调阈值,而应把 MV budget 放进 diffusion 生成过程。

实验结果(表格)

10-case 方法汇总

methodmean ΔChamfermean ΔF@5active ΔChamferactive ΔF@5watertightfacesselected iters
carrier_all0.00000000.0000000.00000000.000000100.0%443226{'0': 10}
a53_saturation_guard-0.00041710.002684-0.00139250.008953100.0%443226{'0': 7, 'None': 3}
a61_min_mean_mvd0.00001660.0010370.00005540.003458100.0%443226{'0': 7, '2': 1, '3': 1, '4': 1}
a61_min_worst_view0.00016900.0007120.00056340.002372100.0%443226{'0': 7, '1': 1, '3': 1, '4': 1}
a61_soft_pareto0.00007140.0004030.00023800.001342100.0%443226{'0': 7, '1': 2, '2': 1}
a61_budgeted_soft_pareto0.00011010.0002020.00036690.000675100.0%443226{'0': 8, '1': 2}
a61_conservative_defer0.00011010.0002020.00036690.000675100.0%443226{'0': 8, '1': 2}
gt_chamfer_oracle-0.00016740.003439-0.00055810.011463100.0%443226{'0': 7, '3': 1, '8': 2}
gt_fscore_oracle-0.00015500.003488-0.00051670.011628100.0%443226{'0': 7, '4': 1, '8': 2}

Active case 选择明细

caseselectoriterΔChamferΔF@5ΔMV mean relΔMV worst relbudget pass
gso_000_input4a61_min_mean_mvd3-0.00041590.000512-0.000990.00326False
gso_000_input4a61_min_worst_view1-0.0002122-0.0002400.000420.00174False
gso_000_input4a61_soft_pareto1-0.0002122-0.0002400.000420.00174False
gso_000_input4a61_budgeted_soft_pareto00.00000000.0000000.000000.00000True
gso_000_input4a61_conservative_defer00.00000000.0000000.000000.00000True
gso_000_input4gt_chamfer_oracle3-0.00041590.000512-0.000990.00326False
gso_002_input4a61_min_mean_mvd4-0.00068180.008366-0.01527-0.00182True
gso_002_input4a61_min_worst_view3-0.00055870.007949-0.01124-0.00404True
gso_002_input4a61_soft_pareto1-0.00033760.002772-0.01164-0.00333True
gso_002_input4a61_budgeted_soft_pareto1-0.00033760.002772-0.01164-0.00333True
gso_002_input4a61_conservative_defer1-0.00033760.002772-0.01164-0.00333True
gso_002_input4gt_chamfer_oracle8-0.00160530.015600-0.01003-0.00392True
gso_008_input4a61_min_mean_mvd20.00126370.001495-0.01031-0.01475True
gso_008_input4a61_min_worst_view40.0024611-0.000593-0.00594-0.01745True
gso_008_input4a61_soft_pareto20.00126370.001495-0.01031-0.01475True
gso_008_input4a61_budgeted_soft_pareto10.0014382-0.000748-0.00864-0.00809True
gso_008_input4a61_conservative_defer10.0014382-0.000748-0.00864-0.00809True
gso_008_input4gt_chamfer_oracle80.00034680.0182770.058470.02036False

MV 信号与 GT Chamfer 的相关性

casesignalSpearman rhop-value
gso_000_input4delta_mv_mean_rel0.46670.2054
gso_000_input4delta_mv_worst_rel0.18330.6368
gso_000_input4soft_mv_score0.18330.6368
gso_000_input4balanced_mv_score0.18330.6368
gso_002_input4delta_mv_mean_rel-0.11670.7650
gso_002_input4delta_mv_worst_rel-0.11670.7650
gso_002_input4soft_mv_score-0.73330.0246
gso_002_input4balanced_mv_score-0.91670.0005
gso_008_input4delta_mv_mean_rel-0.03330.9322
gso_008_input4delta_mv_worst_rel-0.31670.4064
gso_008_input4soft_mv_score-0.15000.7001
gso_008_input4balanced_mv_score0.21670.5755

可视化结果

A61 Chamfer delta 对比
A61 Chamfer delta 对比
A61 F@5 gain 对比
A61 F@5 gain 对比
active cases Chamfer trajectory
active cases Chamfer trajectory
active cases soft MV score trajectory
active cases soft MV score trajectory

实验结论

下一步想法