实验定位 负结果 / 机制有效
A44 试图把 A42 的 component prior 往 signed/local 方向推进：不再用 unsigned UDF 重新抽壳，而是把 A42 单组件水密 mesh 当作 signed/topology carrier，保持面片连接不变，只把顶点向 visible-kept 和 replacement patch target surface 小步投影。
结果是一个很有价值的负结果：水密率保持 100.0%、components 保持 1，visible p95 从 carrier 的 0.03251 降到 0.02730；但 GT Chamfer 从 0.08930 变差到 0.09004。visible-heavy 进一步降低 proxy 到 0.02389，GT 仍更差。
实验设计（Image-2 风格）
模块设计（Image-2 风格）
Signed Carrier
- 输入：A42 visible-prior 单组件水密 mesh。
- 拓扑：不改变 faces/connectivity，carrier 提供闭合符号结构。
- 目标：visible-kept surface 和 raw replacement patch samples。
- 变体：soft、visible-heavy、projection+smooth。
Projection Gate
- nearest target 按 visible/patch label 取不同 alpha。
- tau 限制只移动靠近 target 的顶点。
- max_step 限制单次位移。
- 评估同时看 proxy p95 与 GT Chamfer，避免观测点云过拟合。
实验结果（表格）
聚合结果
| 变体 | GT Chamfer↓ | F@5↑ | 水密率↑ | components↓ | boundary↓ | nonmanifold↓ | visible p95↓ | patch p95↓ | active vertices↓ | p95 move | 结论 |
|---|---|---|---|---|---|---|---|---|---|---|---|
| A42 carrier control | 0.08930 | 0.4495 | 100.0% | 1 | 0 | 0 | 0.03251 | 0.03537 | 0.0% | 0.00000 | carrier baseline |
| soft projection | 0.09004 | 0.4471 | 100.0% | 1 | 0 | 0 | 0.02730 | 0.03070 | 100.0% | 0.00700 | proxy 改善 / GT 变差 |
| visible-heavy projection | 0.09036 | 0.4454 | 100.0% | 1 | 0 | 0 | 0.02389 | 0.02810 | 100.0% | 0.01375 | proxy 改善 / GT 变差 |
| projection + Taubin smooth | 0.09033 | 0.4456 | 100.0% | 1 | 0 | 0 | 0.02635 | 0.03062 | 100.0% | 0.00900 | proxy 改善 / GT 变差 |
逐 case 输出
| case | 变体 | GT Chamfer↓ | F@5↑ | watertight | components | visible p95 | patch p95 | active vertices | p95 move | OBJ |
|---|---|---|---|---|---|---|---|---|---|---|
| gso_002_input4 | A42 carrier control | 0.04584 | 0.6875 | True | 1 | 0.03323 | 0.03794 | 0.0% | 0.00000 | /home/jiachen/TRELLIS/experiments/nonorthogonal_gpt_orthoview/outputs/data_survey/vggt_spatial_adapter/support_guidance_a44_topology_preserving_projection/gso_002_input4/mesh_carrier_control.obj |
| gso_002_input4 | soft projection | 0.04598 | 0.6899 | True | 1 | 0.02884 | 0.03325 | 100.0% | 0.00700 | /home/jiachen/TRELLIS/experiments/nonorthogonal_gpt_orthoview/outputs/data_survey/vggt_spatial_adapter/support_guidance_a44_topology_preserving_projection/gso_002_input4/mesh_signed_project_soft.obj |
| gso_002_input4 | visible-heavy projection | 0.04631 | 0.6878 | True | 1 | 0.02646 | 0.03037 | 100.0% | 0.01375 | /home/jiachen/TRELLIS/experiments/nonorthogonal_gpt_orthoview/outputs/data_survey/vggt_spatial_adapter/support_guidance_a44_topology_preserving_projection/gso_002_input4/mesh_signed_project_visible.obj |
| gso_002_input4 | projection + Taubin smooth | 0.04641 | 0.6872 | True | 1 | 0.02823 | 0.03277 | 100.0% | 0.00900 | /home/jiachen/TRELLIS/experiments/nonorthogonal_gpt_orthoview/outputs/data_survey/vggt_spatial_adapter/support_guidance_a44_topology_preserving_projection/gso_002_input4/mesh_signed_project_smooth.obj |
| gso_008_input4 | A42 carrier control | 0.13276 | 0.2114 | True | 1 | 0.03178 | 0.03280 | 0.0% | 0.00000 | /home/jiachen/TRELLIS/experiments/nonorthogonal_gpt_orthoview/outputs/data_survey/vggt_spatial_adapter/support_guidance_a44_topology_preserving_projection/gso_008_input4/mesh_carrier_control.obj |
| gso_008_input4 | soft projection | 0.13409 | 0.2044 | True | 1 | 0.02576 | 0.02816 | 100.0% | 0.00700 | /home/jiachen/TRELLIS/experiments/nonorthogonal_gpt_orthoview/outputs/data_survey/vggt_spatial_adapter/support_guidance_a44_topology_preserving_projection/gso_008_input4/mesh_signed_project_soft.obj |
| gso_008_input4 | visible-heavy projection | 0.13442 | 0.2029 | True | 1 | 0.02132 | 0.02583 | 100.0% | 0.01375 | /home/jiachen/TRELLIS/experiments/nonorthogonal_gpt_orthoview/outputs/data_survey/vggt_spatial_adapter/support_guidance_a44_topology_preserving_projection/gso_008_input4/mesh_signed_project_visible.obj |
| gso_008_input4 | projection + Taubin smooth | 0.13425 | 0.2040 | True | 1 | 0.02446 | 0.02848 | 100.0% | 0.00900 | /home/jiachen/TRELLIS/experiments/nonorthogonal_gpt_orthoview/outputs/data_survey/vggt_spatial_adapter/support_guidance_a44_topology_preserving_projection/gso_008_input4/mesh_signed_project_smooth.obj |
可视化结果
实验结论
- A44 证明 topology-preserving deformation 是可执行的：所有变体保持 watertight=100%、boundary=0、nonmanifold=0、components=1。
- 但 A44 也证伪了“全局把 carrier 拉向 target 点云”的路线：proxy p95 明显下降，GT Chamfer 和 F@5 反而变差。
- 关键失败原因是 active vertex ratio 接近 100.0%，说明当前 trust region 太宽，几乎全表面都在动，不是 conflict-local edit。
- 这比 A42 更进一步：signed/topology carrier 是正确结构，但 deformation 必须被 A35/A30 的 conflict mask 限制在局部区域，并且可见可信面应作为硬锁而不是 target attractor。
- 这个负结果对论文有价值：它说明 observable-state guidance 的核心不是“向观测点云靠近”，而是“只在可见冲突区域施加有符号残差，同时冻结非冲突可见表面”。
下一步想法
- A45：Conflict-local signed projection，只允许 A35/A30 surface_conflict 附近的 carrier 顶点移动，visible-kept 区域作为 hard lock。
- A46：Signed residual SDF，在局部 tube 内用 target residual 改变 signed distance，而不是直接 vertex projection。
- A47：把 active vertex ratio 作为新 gate：任何无 GT 修复如果激活面积过大，应自动降级为 defer。
- A48：扩展到 10-case，验证“proxy 改善但 GT 变差”的过拟合现象是否普遍。