实验定位 负结果 / 有指导价值
A20 接着 A19 的 hard case gso_008_input4 往下追：如果 whole-mesh repair 因 under-coverage 被拒绝，能不能只在局部使用 repair，同时保留 baseline 的可见轮廓？
结论很清楚：对 gso_008，repair 没有任何输入视角 IoU 优于 baseline，并且会删除输入可见区域。这个负结果说明 naive 局部拼接也不可靠，下一步必须把“可见支持场”或 depth-aware completion 放到生成/提取阶段。
实验设计（Image-2 风格）
模块设计（Image-2 风格）
Gate Purpose
Reject global repair when it erases visible baseline support; only future local repair candidates with positive per-view local margin should be allowed.
This is an analysis gate, not a generated hybrid mesh. It estimates whether a future component-level 3D operation is justified.
Rules
- repair must be watertight and topology-clean
- no view may lose more than 8% target-visible baseline support
- mean visible loss ratio must be <= 4%
- repair must improve at least one input-view IoU by > 0.01
- mean local margin must be >= -1%
实验结果（表格）
| case | local gate | A19选择 | repair IoU Δ | repair更好视角数 | 可见损失 | safe add | local margin | GT Chamfer Δ | GT F@5 Δ | 拒绝原因 |
|---|---|---|---|---|---|---|---|---|---|---|
| gso_000_input4 | pass | ranker_selected | 0.1400 | 4 | 0.0338 | 0.3528 | 0.3191 | -0.00160 | 0.0202 | [] |
| gso_001_input4 | reject | ranker_selected | -0.0013 | 0 | 0.0061 | 0.0049 | -0.0012 | -0.00105 | 0.0082 | ['repair is not better than baseline in any visible view'] |
| gso_002_input4 | pass | ranker_selected | 0.0228 | 3 | 0.0226 | 0.0570 | 0.0344 | -0.00377 | 0.0502 | [] |
| gso_003_input4 | reject | ranker_selected | -0.0001 | 0 | 0.0014 | 0.0000 | -0.0014 | -0.00005 | -0.0041 | ['repair is not better than baseline in any visible view'] |
| gso_004_input4 | reject | ranker_selected | -0.0006 | 0 | 0.0005 | 0.0004 | -0.0001 | -0.00041 | 0.0032 | ['repair is not better than baseline in any visible view'] |
| gso_005_input4 | reject | ranker_selected | 0.0000 | 0 | 0.0000 | 0.0000 | 0.0000 | 0.00119 | -0.0079 | ['repair is not better than baseline in any visible view'] |
| gso_006_input4 | reject | repair_hybrid | -0.0831 | 0 | 0.0217 | 0.2262 | 0.2045 | 0.00144 | 0.0204 | ['repair is not better than baseline in any visible view'] |
| gso_007_input4 | reject | ranker_selected | 0.0013 | 0 | 0.0031 | 0.0074 | 0.0044 | 0.00281 | -0.0188 | ['repair is not better than baseline in any visible view'] |
| gso_008_input4 | reject | baseline | -0.0237 | 0 | 0.0363 | 0.0014 | -0.0349 | 0.00474 | -0.0168 | ['repair deletes visible baseline support in at least one view', 'repair is not better than baseline in any visible view', 'local silhouette margin is negative'] |
| gso_009_input4 | pass | ranker_selected | -0.0027 | 1 | 0.0339 | 0.0311 | -0.0028 | -0.00006 | -0.0029 | [] |
gso_008 逐视角诊断
| 视角 | baseline IoU | repair IoU | repair IoU Δ | 可见损失 | safe add | local margin |
|---|---|---|---|---|---|---|
| view 0 | 0.2494 | 0.2362 | -0.0131 | 0.0308 | 0.0000 | -0.0308 |
| view 1 | 0.2072 | 0.2025 | -0.0047 | 0.0066 | 0.0010 | -0.0056 |
| view 2 | 0.3419 | 0.2806 | -0.0613 | 0.0893 | 0.0012 | -0.0881 |
| view 3 | 0.1915 | 0.1759 | -0.0156 | 0.0186 | 0.0035 | -0.0151 |
可视化结果
gso_008 可见覆盖 overlay
这些图用于判断 repair 是否在输入可见区域“抹掉”了 baseline 已经覆盖的结构。对 gso_008，关键视角出现明显蓝色区域，代表 repair 删除了 baseline 的可见支持。
实验结论
- 10 个 case 中 local gate 通过 3 个，但 A19 仍 fallback baseline 的 hard case 通过数为 0。
- gso_008 的 repair 更好视角数为 0，最大可见损失为 0.0893，mean local margin=-0.0349。
- 事后 GT 也支持拒绝：gso_008 repair 相比 baseline Chamfer Δ=0.00474，F@5 Δ=-0.0168。
- 所以 A20 没有产生更好的 mesh，但它给出一个重要限制：水密候选不能只因 topology 干净就参与局部融合，必须证明它不删除输入可见支持。
- 这把下一步创新点从“mesh repair/routing”推进到“view-conditioned support field”：先在 3D token/voxel/SLat 层标记哪些区域由输入视角支持，再做闭合和减面。
下一步想法
- A21：把 A20 的 2D visible-safe 证据提升到 3D support field：用输入相机反投影 silhouette/depth，把 sparse token 标成 visible-supported / unobserved / conflict。
- 只允许 watertight closure 改写 unobserved 或 low-confidence 区域；visible-supported 区域用 TRELLIS.2 baseline 或 VGGT-consistent token anchor 锁住。
- 引入 depth/normal reprojection，而不是只用 silhouette；这样能避免外轮廓相同但厚度、凹凸、背面结构错误的问题。
- 对 100-case 跑 A19/A20 taxonomy 统计，判断 under-coverage 是个别 case 还是主要失败模式。