实验定位 10-case mask artifact 正向
A50 接在 A49 后面：A49 已经把 VGGT depth residual 转成 10-case sample-level surface labels，但 sample 点云还不能直接驱动 mesh 局部编辑。A50 把这些标签投票到 repair mesh 的 faces 和 vertices，生成后续 carrier、local residual、hard-lock、defer routing 都能直接读取的 editable mask。
结果：requested=10，completed=10，skipped=0；mask NPZ、face PLY、vertex PLY 都持久化 10/10。平均 face confidence=0.9404，vertex confidence=0.9758，p95 transfer distance=0.01187。
实验设计（Image-2 风格）
模块设计（Image-2 风格）
Voting Interface
- face label：每个 face centroid 查询最近 5 个 A49 surface samples。
- 投票规则：多数标签为主，conflict/improved 等标签用固定 priority 打破平票。
- vertex label： incident faces 按 face confidence 加权投票。
- 输出：
face_label_ids、face_confidence、vertex_label_ids、vertex_confidence。
Research Meaning
A50 的作用是把“观测状态证据”变成“mesh 可执行控制层”。它本身不编辑 geometry，因此不报告 Chamfer 提升；真正价值在于为 A47/A48 的 10-case rerun 提供局部活动区、hard-lock 区和 defer 区。
实验结果（表格）
| case | faces | vertices | face conf | vertex conf | p95 transfer dist | improved face | conflict face | neutral face | unobserved face | mask NPZ | face PLY |
|---|---|---|---|---|---|---|---|---|---|---|---|
| gso_000_input4 | 2734 | 1369 | 0.9110 | 0.8494 | 0.01337 | 0.4952 | 0.1306 | 0.1383 | 0.2359 | /home/jiachen/TRELLIS/experiments/nonorthogonal_gpt_orthoview/outputs/data_survey/vggt_spatial_adapter/support_guidance_a50_10case_face_vertex_mask_expansion/gso_000_input4/face_vertex_masks.npz | /home/jiachen/TRELLIS/experiments/nonorthogonal_gpt_orthoview/outputs/data_survey/vggt_spatial_adapter/support_guidance_a50_10case_face_vertex_mask_expansion/gso_000_input4/mesh_face_mask.ply |
| gso_001_input4 | 449604 | 224802 | 0.9697 | 0.9946 | 0.01112 | 0.0156 | 0.0290 | 0.9554 | 0.0000 | /home/jiachen/TRELLIS/experiments/nonorthogonal_gpt_orthoview/outputs/data_survey/vggt_spatial_adapter/support_guidance_a50_10case_face_vertex_mask_expansion/gso_001_input4/face_vertex_masks.npz | /home/jiachen/TRELLIS/experiments/nonorthogonal_gpt_orthoview/outputs/data_survey/vggt_spatial_adapter/support_guidance_a50_10case_face_vertex_mask_expansion/gso_001_input4/mesh_face_mask.ply |
| gso_002_input4 | 234960 | 117482 | 0.9010 | 0.9796 | 0.01703 | 0.6134 | 0.2018 | 0.1583 | 0.0265 | /home/jiachen/TRELLIS/experiments/nonorthogonal_gpt_orthoview/outputs/data_survey/vggt_spatial_adapter/support_guidance_a50_10case_face_vertex_mask_expansion/gso_002_input4/face_vertex_masks.npz | /home/jiachen/TRELLIS/experiments/nonorthogonal_gpt_orthoview/outputs/data_survey/vggt_spatial_adapter/support_guidance_a50_10case_face_vertex_mask_expansion/gso_002_input4/mesh_face_mask.ply |
| gso_003_input4 | 295076 | 147538 | 0.8489 | 0.9701 | 0.01880 | 0.0895 | 0.0897 | 0.8208 | 0.0000 | /home/jiachen/TRELLIS/experiments/nonorthogonal_gpt_orthoview/outputs/data_survey/vggt_spatial_adapter/support_guidance_a50_10case_face_vertex_mask_expansion/gso_003_input4/face_vertex_masks.npz | /home/jiachen/TRELLIS/experiments/nonorthogonal_gpt_orthoview/outputs/data_survey/vggt_spatial_adapter/support_guidance_a50_10case_face_vertex_mask_expansion/gso_003_input4/mesh_face_mask.ply |
| gso_004_input4 | 100120 | 50062 | 0.9895 | 0.9973 | 0.01353 | 0.0038 | 0.0038 | 0.9925 | 0.0000 | /home/jiachen/TRELLIS/experiments/nonorthogonal_gpt_orthoview/outputs/data_survey/vggt_spatial_adapter/support_guidance_a50_10case_face_vertex_mask_expansion/gso_004_input4/face_vertex_masks.npz | /home/jiachen/TRELLIS/experiments/nonorthogonal_gpt_orthoview/outputs/data_survey/vggt_spatial_adapter/support_guidance_a50_10case_face_vertex_mask_expansion/gso_004_input4/mesh_face_mask.ply |
| gso_005_input4 | 328064 | 164030 | 1.0000 | 1.0000 | 0.00589 | 0.0000 | 0.0000 | 1.0000 | 0.0000 | /home/jiachen/TRELLIS/experiments/nonorthogonal_gpt_orthoview/outputs/data_survey/vggt_spatial_adapter/support_guidance_a50_10case_face_vertex_mask_expansion/gso_005_input4/face_vertex_masks.npz | /home/jiachen/TRELLIS/experiments/nonorthogonal_gpt_orthoview/outputs/data_survey/vggt_spatial_adapter/support_guidance_a50_10case_face_vertex_mask_expansion/gso_005_input4/mesh_face_mask.ply |
| gso_006_input4 | 1832108 | 911490 | 0.8698 | 0.9848 | 0.01430 | 0.2711 | 0.4054 | 0.3225 | 0.0010 | /home/jiachen/TRELLIS/experiments/nonorthogonal_gpt_orthoview/outputs/data_survey/vggt_spatial_adapter/support_guidance_a50_10case_face_vertex_mask_expansion/gso_006_input4/face_vertex_masks.npz | /home/jiachen/TRELLIS/experiments/nonorthogonal_gpt_orthoview/outputs/data_survey/vggt_spatial_adapter/support_guidance_a50_10case_face_vertex_mask_expansion/gso_006_input4/mesh_face_mask.ply |
| gso_007_input4 | 278432 | 139218 | 0.9925 | 0.9984 | 0.00771 | 0.0014 | 0.0264 | 0.9722 | 0.0000 | /home/jiachen/TRELLIS/experiments/nonorthogonal_gpt_orthoview/outputs/data_survey/vggt_spatial_adapter/support_guidance_a50_10case_face_vertex_mask_expansion/gso_007_input4/face_vertex_masks.npz | /home/jiachen/TRELLIS/experiments/nonorthogonal_gpt_orthoview/outputs/data_survey/vggt_spatial_adapter/support_guidance_a50_10case_face_vertex_mask_expansion/gso_007_input4/mesh_face_mask.ply |
| gso_008_input4 | 827994 | 413999 | 0.9431 | 0.9913 | 0.00802 | 0.1707 | 0.7050 | 0.1241 | 0.0002 | /home/jiachen/TRELLIS/experiments/nonorthogonal_gpt_orthoview/outputs/data_survey/vggt_spatial_adapter/support_guidance_a50_10case_face_vertex_mask_expansion/gso_008_input4/face_vertex_masks.npz | /home/jiachen/TRELLIS/experiments/nonorthogonal_gpt_orthoview/outputs/data_survey/vggt_spatial_adapter/support_guidance_a50_10case_face_vertex_mask_expansion/gso_008_input4/mesh_face_mask.ply |
| gso_009_input4 | 83166 | 41585 | 0.9784 | 0.9922 | 0.00889 | 0.0205 | 0.0176 | 0.9619 | 0.0000 | /home/jiachen/TRELLIS/experiments/nonorthogonal_gpt_orthoview/outputs/data_survey/vggt_spatial_adapter/support_guidance_a50_10case_face_vertex_mask_expansion/gso_009_input4/face_vertex_masks.npz | /home/jiachen/TRELLIS/experiments/nonorthogonal_gpt_orthoview/outputs/data_survey/vggt_spatial_adapter/support_guidance_a50_10case_face_vertex_mask_expansion/gso_009_input4/mesh_face_mask.ply |
可视化结果
实验结论
- A50 成功把 A49 的 10-case sample labels 全部投票到 repair mesh faces/vertices：completed=10/10，skipped=0。
- 平均 face confidence=0.9404、vertex confidence=0.9758，说明 sample-to-face 标签传递整体稳定；p95 nearest-sample distance=0.01187，没有出现明显离面投票。
- 平均 conflict face ratio=16.1%，与 A49 sample conflict ratio 接近，说明 face mask 没有把局部冲突大量稀释掉。
- gso_006/gso_008 这类高 conflict case 可以进入 10-case depth-sign residual；高 neutral case 则应由 A48-style confidence gate defer，不应强行移动。
- 限制：A50 只是 mask artifact，不是最终 mesh edit；下一步必须生成 10-case topology carrier，并在 A50 mask 上重跑 A47/A48，才能判断几何质量是否真的提升。
下一步想法
- A51：扩展 A42 component-aware topology carrier 到 A31 10-case，并把 A50 face/vertex masks 对齐到 carrier mesh。
- A52：在 10-case 上重跑 A47 depth-sign residual，比较 carrier / signed / outward / inward / smooth。
- A53：重跑 A48 no-GT confidence gate，验证 2-case selector 是否能扩展到 10-case。
- A54：引入 mask-aware Laplacian regularization：conflict faces 活动、visible/improved faces hard-lock、neutral faces soft-lock。