实验定位 诊断正结果 / 路由混合
A27 回答 A26 的核心问题：为什么 case-level risk 分不开某些 under-coverage？本轮把 A20 的 per-view visible evidence 和 A23 的 per-view VGGT depth/normal evidence 对齐，给每个输入视角打标签：depth_good、depth_bad、visible_conflict、silhouette_good、under_coverage。
焦点结论很清楚：gso_002 虽然有 under-coverage，但有 depth_good views=[0, 1] 和 silhouette_good views=[3]，所以 repair 可接受；gso_008 同时有 case/per-view depth reject、visible_conflict views=[2] 和 silhouette_bad，因此必须拒绝。
实验设计（Image-2 风格）
模块设计（Image-2 风格）
Hypothesis
问题：A26 case-level risk cannot distinguish local depth-supported under-coverage from true visible support deletion.
Per-view lock：Align per-view VGGT depth deltas with per-view visible loss and silhouette deltas to classify each view as depth_good/depth_bad/visible_conflict/etc.
论文角度：This is the first step toward a pixel/region-level lock map that separates visible-locked, unobserved-closable and conflict regions.
Taxonomy Tags
- depth_good/depth_bad：repair 相对 baseline 是否更接近 VGGT depth。
- visible_conflict：repair 删除可见支持且 silhouette 变差。
- silhouette_good：repair 在该视角改善轮廓，可抵消轻微 coverage 风险。
- under_coverage：区域缩小，只能作为风险提示，不能单独拒绝。
实验结果（表格）
| 策略 | 选择 repair 数 | false reject | false accept | selected Chamfer Δ | selected F@5 Δ | 选择的 repair case |
|---|---|---|---|---|---|---|
| A24 calibrated | 3 | 4 | 0 | -0.00054 | 0.0067 | gso_000_input4, gso_002_input4, gso_009_input4 |
| A27 per-view depth lock | 5 | 3 | 1 | -0.00030 | 0.0051 | gso_000_input4, gso_002_input4, gso_003_input4, gso_004_input4, gso_007_input4 |
A27 逐 case 汇总
| case | A27 gate | mean view score | depth good views | depth bad views | visible conflict views | silhouette good views | under-coverage views | GT Chamfer Δ | GT F@5 Δ | 原因 |
|---|---|---|---|---|---|---|---|---|---|---|
| gso_000_input4 | pass | 0.1363 | [0, 1, 2, 3] | [] | [] | [0, 1, 2, 3] | [] | -0.00160 | 0.0202 | [] |
| gso_001_input4 | reject | -0.0019 | [] | [] | [] | [] | [0, 1, 2, 3] | -0.00105 | 0.0082 | ['broad under-coverage without depth/silhouette compensation'] |
| gso_002_input4 | pass | 0.0152 | [0, 1] | [2] | [] | [3] | [1, 2, 3] | -0.00377 | 0.0502 | [] |
| gso_003_input4 | pass | -0.0027 | [] | [] | [] | [] | [] | -0.00005 | -0.0041 | [] |
| gso_004_input4 | pass | -0.0003 | [] | [] | [] | [] | [] | -0.00041 | 0.0032 | [] |
| gso_005_input4 | reject | -0.0000 | [] | [] | [] | [] | [0, 1, 2, 3] | 0.00119 | -0.0079 | ['broad under-coverage without depth/silhouette compensation'] |
| gso_006_input4 | reject | -0.0255 | [] | [0, 1, 2, 3] | [] | [] | [] | 0.00144 | 0.0204 | ['case and per-view depth both reject', 'multi-view depth_bad'] |
| gso_007_input4 | pass | -0.0001 | [] | [3] | [] | [] | [] | 0.00281 | -0.0188 | [] |
| gso_008_input4 | reject | -0.0382 | [2] | [0] | [2] | [] | [0, 1, 2, 3] | 0.00474 | -0.0168 | ['case and per-view depth both reject', 'visible conflict without compensating silhouette'] |
| gso_009_input4 | reject | -0.0084 | [] | [] | [0] | [1] | [2] | -0.00006 | -0.0029 | [] |
gso_002 vs gso_008 逐视角解释
| case | view | score | depth Δ | normal Δ | visible loss | IoU Δ | area ratio | tags |
|---|---|---|---|---|---|---|---|---|
| gso_002_input4 | 0 | 0.0645 | -0.05570 | -0.0658 | 0.0178 | 0.0152 | 0.979 | ['depth_good'] |
| gso_002_input4 | 1 | 0.0300 | -0.02306 | -0.0693 | 0.0180 | 0.0189 | 0.727 | ['depth_good', 'under_coverage'] |
| gso_002_input4 | 2 | -0.0563 | 0.05237 | -0.0784 | 0.0386 | 0.0079 | 0.618 | ['depth_bad', 'under_coverage'] |
| gso_002_input4 | 3 | 0.0224 | 0.00016 | -0.0741 | 0.0160 | 0.0492 | 0.732 | ['silhouette_good', 'under_coverage'] |
| gso_008_input4 | 0 | -0.0878 | 0.07355 | 0.0839 | 0.0308 | -0.0131 | 0.685 | ['depth_bad', 'under_coverage'] |
| gso_008_input4 | 1 | -0.0078 | 0.00355 | 0.0013 | 0.0066 | -0.0047 | 0.207 | ['under_coverage'] |
| gso_008_input4 | 2 | -0.0437 | -0.01672 | -0.0095 | 0.0893 | -0.0613 | 0.635 | ['depth_good', 'visible_conflict', 'silhouette_bad', 'under_coverage'] |
| gso_008_input4 | 3 | -0.0136 | 0.00040 | -0.0061 | 0.0186 | -0.0156 | 0.202 | ['under_coverage'] |
可视化结果
实验结论
- A27 的解释能力明显强于 A26：它能说明 gso_002 与 gso_008 的差异不是“是否 under-coverage”，而是 under-coverage 是否被 depth_good / silhouette_good 支持，或是否伴随 visible_conflict。
- 作为路由器，A27 仍然混合：repair selected=5，false reject=3，false accept=1；A24 仍然更稳但更保守。
- A27 相比 A24 减少一个 false reject，但引入 gso_007 false accept，说明 view-level taxonomy 还需要从 view 级升级到 pixel/region 级。
- 论文主线更清楚了：最终贡献不应是一个 scalar gate，而应是 view-conditioned observable-region locking，把可见支持、depth consistency 和未观测闭合分开建模。
- 这一步已经具备可发表方法雏形的分析证据：非正交多视角输入的失败来自局部可观测性冲突，而不是单纯视角数量或全局融合权重不足。
下一步想法
- A28：把 view-level taxonomy 细化到 pixel/region lock map，输出 visible-locked / depth-conflict / silhouette-only / unobserved-closable 四类区域图。
- 给 gso_002、gso_008、gso_007 做可视化面板：每个输入视角显示 depth Δ、visible loss、IoU Δ 和最终标签，验证标签是否符合人眼判断。
- 将 lock map 投影到 mesh surface：只允许 repair 改写 unobserved-closable 区域，visible-locked 区域必须保留 baseline 或 VGGT-consistent surface。
- 把 lock map 作为 diffusion sampler 的局部约束，而不是 post-hoc 选择器，这是更接近论文贡献的实现路线。