实验 31:A25 Regional Visible Geometry Gate

把整 case visible-safe 拆成 per-view / per-region 证据,检验区域可见约束的边界

实验定位 负结果 / 关键诊断

A25 追问 A24 的遗留问题:A20 的 visible-safe gate 是整 case 拒绝,明显过保守;如果把可见区域证据拆成 per-view/per-region,是否能减少误拒?本实验仍是无 GT 路由,GT 只用于事后标注 false reject / false accept。

结果很有信息量:A25 regional 把 false reject 从 A24 的 4 降到 2,但 false accept 从 0 增到 2,放过了 gso_005_input4, gso_007_input4。depth-coupled 变体消除了 false accept,但 false reject 升到 5,说明区域 silhouette 本身不足,需要更可靠的 depth-confidence 局部约束。

实验设计(Image-2 风格)

实验设计图:A25 regional visible geometry gate
实验设计图:A25 regional visible geometry gate
模块设计图:regional visible support taxonomy
模块设计图:regional visible support taxonomy

模块设计(Image-2 风格)

Hypothesis

问题:A20 case-level visible-safe veto rejects any case with visible risk, which is too conservative for small single-view or repair-compensated deviations.

区域门控:Use per-view visible loss, IoU delta and under-coverage taxonomy; reject only multi-view severe support deletion or deletion without depth/IoU compensation.

论文角度:Move from scalar candidate routing to view-conditioned visible support constraints, a prerequisite for generation-time local locking.

Decision Signals

  • per-view visible loss:repair 是否删除输入可见支持。
  • per-view IoU delta:repair 相对 baseline 是否改善某些视角。
  • under/over-coverage taxonomy:区分轮廓缩小、扩张和中性变化。
  • VGGT depth delta:阻止 silhouette-only gate 放行几何坏修复。

实验结果(表格)

策略选择 repair 数false rejectfalse acceptselected Chamfer Δselected F@5 Δ选择的 repair case
A20 case visible-safe340-0.000540.0067gso_000_input4, gso_002_input4, gso_009_input4
A24 calibrated340-0.000540.0067gso_000_input4, gso_002_input4, gso_009_input4
A25 regional visible722-0.000290.0051gso_000_input4, gso_001_input4, gso_002_input4, gso_003_input4, gso_004_input4, gso_005_input4, gso_007_input4
A25 depth-coupled regional250-0.000540.0070gso_000_input4, gso_002_input4

A25 regional 逐 case 决策

caseA25 gatescoreVGGT depth Δmean IoU Δmax visible losssevere loss viewsunder-coverage viewsGT Chamfer ΔGT F@5 Δ原因
gso_000_input4pass0.2633-0.023300.14000.0538[][]-0.001600.0202[]
gso_001_input4pass0.0002-0.00020-0.00130.0080[][0, 1, 2, 3]-0.001050.0082[]
gso_002_input4pass0.0666-0.006560.02280.0386[][1, 2, 3]-0.003770.0502[]
gso_003_input4pass-0.00190.00193-0.00010.0022[][]-0.00005-0.0041[]
gso_004_input4pass0.0001-0.00009-0.00060.0020[][1, 3]-0.000410.0032[]
gso_005_input4pass-0.00010.000080.00000.0000[][0, 1, 2, 3]0.00119-0.0079[]
gso_006_input4reject-0.39040.03191-0.08310.0477[][3]0.001440.0204['VGGT depth hard reject']
gso_007_input4pass-0.00140.001940.00130.0086[][1, 3]0.00281-0.0188[]
gso_008_input4reject-0.22480.01519-0.02370.0893[2][0, 1, 2, 3]0.00474-0.0168['VGGT depth hard reject', 'severe visible loss with negative mean IoU delta', 'multi-view under-coverage without depth compensation']
gso_009_input4reject-0.03010.00017-0.00270.0722[][0, 2, 3]-0.00006-0.0029['multi-view under-coverage without depth compensation']

可视化结果

A25 路由错误类型对比
A25 路由错误类型对比
A25 接受 repair 数和 F@5 收益对比
A25 接受 repair 数和 F@5 收益对比

实验结论

下一步想法