实验 29:A23 VGGT Depth/Normal Gate

用 VGGT 预测几何替代 GT oracle,测试无 GT depth/normal 可见一致性门控

实验定位 混合 / 方向有效但需校准

A23 接着 A22 做了关键替换:A22 用 GT mesh 做 depth/normal oracle,只能证明信号价值;A23 改成只使用输入图像和 VGGT 预测的 world points、confidence、相对 depth,测试一个真正无 GT 的 depth/normal repair gate。

结果不是简单正结果。VGGT depth 能抓住坏修复,例如 gso_008 的 repair depth error 增加 0.01519,GT Chamfer 也变差 0.00474;但 VGGT normal 由 world_points 有限差分得到,噪声会导致误拒,例如 gso_002 被 normal 拒绝,实际 GT Chamfer/F@5 都改善。

实验设计(Image-2 风格)

实验设计图:VGGT-predicted depth/normal gate
实验设计图:VGGT-predicted depth/normal gate
模块设计图:无 GT deployable geometry gate
模块设计图:无 GT deployable geometry gate

模块设计(Image-2 风格)

Gate Rules

  • compare candidate rendered depth to VGGT depth after robust per-view scale alignment
  • compare candidate rendered normals to normals estimated from VGGT world_points
  • use VGGT confidence quantile to keep high-confidence pixels
  • reject repair when VGGT depth/normal worsens or A20 rejects visible safety

Protocol / Limitations

case:10 个,输入视角 4 张,VGGT 分辨率 96。

  • VGGT depth scale is arbitrary, so depth is median-scale aligned per view
  • VGGT normals are finite-difference estimates from predicted world points
  • this tests routing signal quality, not a new generated mesh yet

实验结果(表格)

方法caseVGGT depth abs ↓VGGT normal cos ↑depth Δnormal Δ
baseline100.049270.4527--
repair_hybrid100.051380.45740.002110.0047

逐 case VGGT gate

caseA23 gatebaseline depthrepair depthdepth Δbaseline normalrepair normalnormal ΔGT Chamfer ΔGT F@5 ΔA20 gateA22 oracle gate原因
gso_000_input4pass0.078230.05493-0.023300.51030.5079-0.0024-0.001600.0202TrueFalse[]
gso_001_input4reject0.036840.03664-0.000200.40450.4013-0.0031-0.001050.0082FalseFalse['A20 2D visible-safe gate rejects repair']
gso_002_input4reject0.088960.08240-0.006560.69440.6225-0.0719-0.003770.0502TrueTrue['repair worsens VGGT normal']
gso_003_input4reject0.025430.027360.001930.54770.5452-0.0025-0.00005-0.0041FalseFalse['A20 2D visible-safe gate rejects repair']
gso_004_input4reject0.036510.03641-0.000090.48840.4865-0.0019-0.000410.0032FalseFalse['A20 2D visible-safe gate rejects repair']
gso_005_input4reject0.007070.007150.000080.25840.25960.00110.00119-0.0079FalseFalse['A20 2D visible-safe gate rejects repair']
gso_006_input4reject0.070580.102480.031910.51580.59330.07750.001440.0204FalseFalse['repair worsens VGGT depth', 'A20 2D visible-safe gate rejects repair']
gso_007_input4reject0.100690.102630.001940.53940.56710.02770.00281-0.0188FalseFalse['A20 2D visible-safe gate rejects repair']
gso_008_input4reject0.031800.047000.015190.23920.25660.01740.00474-0.0168FalseFalse['repair worsens VGGT depth', 'A20 2D visible-safe gate rejects repair']
gso_009_input4pass0.016580.016750.000170.32840.33360.0052-0.00006-0.0029TrueTrue[]

可视化结果

A23 VGGT depth/normal aggregate metrics
A23 VGGT depth/normal aggregate metrics
A23 repair vs baseline VGGT deltas
A23 repair vs baseline VGGT deltas

gso_008 VGGT depth / normal panels

gso_008 baseline:VGGT depth vs candidate depth
gso_008 baseline:VGGT depth vs candidate depth
gso_008 repair:VGGT depth vs candidate depth
gso_008 repair:VGGT depth vs candidate depth
gso_008 baseline:VGGT normal vs candidate normal
gso_008 baseline:VGGT normal vs candidate normal
gso_008 repair:VGGT normal vs candidate normal
gso_008 repair:VGGT normal vs candidate normal

实验结论

下一步想法