实验定位 混合诊断 / 关键验证
A43 是 A42 之后必须做的“回到 GT”验证：A42 在局部 proxy 上看起来很好，但只有 GT Chamfer/F-score 才能证明它没有把形状修坏。这里比较 TRELLIS raw、TRELLIS.2 raw、A37 voxel closure、A40 local UDF、A41 visible-lock UDF 和 A42 component cleanup。
结论非常清楚：A42 相比 TRELLIS.2 raw 把水密率从 0.0% 提到 100.0%，components 从 2328.5 降到 1.0，F@5 从 0.4023 提到 0.4531；但 TRELLIS raw 的 GT Chamfer=0.05540 仍明显优于 A42 的 0.08827。
实验设计（Image-2 风格）
模块设计（Image-2 风格）
Validation Inputs
- case：['gso_002_input4', 'gso_008_input4']。
- normalization：bbox diagonal normalization, same family as evaluate_benchmark_methods.py。
- surface samples：20000。
- retrospective uses GT：True。
Metrics
- 几何：Chamfer mean/p95、precision、recall、F-score@2/5。
- 拓扑：watertight、boundary edges、nonmanifold edges、connected components。
- 复杂度：faces、vertices。
- 判读：先看 GT 几何，再看 topology/face budget，避免单指标结论。
实验结果（表格）
聚合结果
| 方法 | Chamfer mean↓ | Chamfer p95↓ | F@2↑ | F@5↑ | 水密率↑ | boundary↓ | nonmanifold↓ | components↓ | faces↓ | 结论 |
|---|---|---|---|---|---|---|---|---|---|---|
| TRELLIS raw | 0.05540 | 0.13458 | 0.3645 | 0.6148 | 100.0% | 0.0 | 0.0 | 20.0 | 777944 | GT 几何最强参照 |
| TRELLIS.2 raw | 0.09230 | 0.20902 | 0.1873 | 0.4023 | 0.0% | 33342.5 | 6131.0 | 2328.5 | 1005781 | 非水密碎片化 |
| A37 voxel close | 0.09825 | 0.22658 | 0.2198 | 0.3961 | 100.0% | 0.0 | 0.0 | 6.0 | 15540 | 水密但几何代价大 |
| A40 local UDF | 0.08753 | 0.21054 | 0.2523 | 0.4543 | 100.0% | 0.0 | 0.0 | 1.0 | 18316 | 修复拓扑 / 几何未胜 TRELLIS |
| A41 visible lock | 0.08804 | 0.21249 | 0.2608 | 0.4543 | 100.0% | 0.0 | 0.0 | 5.5 | 28608 | 修复拓扑 / 几何未胜 TRELLIS |
| A42 component cleanup | 0.08827 | 0.21265 | 0.2623 | 0.4531 | 100.0% | 0.0 | 0.0 | 1.0 | 28460 | 拓扑有效 / GT 几何混合 |
逐 case 输出
| case | 方法 | Chamfer mean↓ | F@5↑ | watertight | boundary | nonmanifold | components | faces | OBJ |
|---|---|---|---|---|---|---|---|---|---|
| gso_002_input4 | TRELLIS raw | 0.02760 | 0.8456 | True | 0 | 0 | 38 | 1266620 | /home/jiachen/TRELLIS/experiments/nonorthogonal_gpt_orthoview/outputs/data_survey/method_eval/trellis_protocol4_all100/gso_002_input4/mesh_raw.obj |
| gso_002_input4 | TRELLIS.2 raw | 0.05407 | 0.5947 | False | 14605 | 10921 | 2955 | 1462466 | /home/jiachen/TRELLIS/experiments/nonorthogonal_gpt_orthoview/outputs/data_survey/method_eval/trellis2_protocol4_all100/gso_002_input4/mesh_raw.obj |
| gso_002_input4 | A37 voxel close | 0.04745 | 0.6699 | True | 0 | 0 | 2 | 27936 | /home/jiachen/TRELLIS/experiments/nonorthogonal_gpt_orthoview/outputs/data_survey/vggt_spatial_adapter/support_guidance_a37_deletion_closure_repair/gso_002_input4/mesh_voxel_close_p025.obj |
| gso_002_input4 | A40 local UDF | 0.04606 | 0.6843 | True | 0 | 0 | 1 | 29420 | /home/jiachen/TRELLIS/experiments/nonorthogonal_gpt_orthoview/outputs/data_survey/vggt_spatial_adapter/support_guidance_a40_local_implicit_completion/gso_002_input4/mesh_udf_pitch025_r035_trellis2_raw.obj |
| gso_002_input4 | A41 visible lock | 0.04565 | 0.6895 | True | 0 | 0 | 9 | 46788 | /home/jiachen/TRELLIS/experiments/nonorthogonal_gpt_orthoview/outputs/data_survey/vggt_spatial_adapter/support_guidance_a41_visible_lock_udf_completion/gso_002_input4/mesh_lock_r018_patch030_b024_trellis2_raw.obj |
| gso_002_input4 | A42 component cleanup | 0.04562 | 0.6906 | True | 0 | 0 | 1 | 46644 | /home/jiachen/TRELLIS/experiments/nonorthogonal_gpt_orthoview/outputs/data_survey/vggt_spatial_adapter/support_guidance_a42_component_aware_cleanup/gso_002_input4/mesh_lock_r018_patch030_b024_visible_prior_trellis2_raw.obj |
| gso_008_input4 | TRELLIS raw | 0.08319 | 0.3841 | True | 0 | 0 | 2 | 289268 | /home/jiachen/TRELLIS/experiments/nonorthogonal_gpt_orthoview/outputs/data_survey/method_eval/trellis_protocol4_all100/gso_008_input4/mesh_raw.obj |
| gso_008_input4 | TRELLIS.2 raw | 0.13053 | 0.2099 | False | 52080 | 1341 | 1702 | 549096 | /home/jiachen/TRELLIS/experiments/nonorthogonal_gpt_orthoview/outputs/data_survey/method_eval/trellis2_protocol4_all100/gso_008_input4/mesh_raw.obj |
| gso_008_input4 | A37 voxel close | 0.14905 | 0.1223 | True | 0 | 0 | 10 | 3144 | /home/jiachen/TRELLIS/experiments/nonorthogonal_gpt_orthoview/outputs/data_survey/vggt_spatial_adapter/support_guidance_a37_deletion_closure_repair/gso_008_input4/mesh_voxel_close_p025.obj |
| gso_008_input4 | A40 local UDF | 0.12899 | 0.2243 | True | 0 | 0 | 1 | 7212 | /home/jiachen/TRELLIS/experiments/nonorthogonal_gpt_orthoview/outputs/data_survey/vggt_spatial_adapter/support_guidance_a40_local_implicit_completion/gso_008_input4/mesh_udf_pitch025_r035_trellis2_raw.obj |
| gso_008_input4 | A41 visible lock | 0.13044 | 0.2191 | True | 0 | 0 | 2 | 10428 | /home/jiachen/TRELLIS/experiments/nonorthogonal_gpt_orthoview/outputs/data_survey/vggt_spatial_adapter/support_guidance_a41_visible_lock_udf_completion/gso_008_input4/mesh_lock_r018_patch030_b024_trellis2_raw.obj |
| gso_008_input4 | A42 component cleanup | 0.13091 | 0.2157 | True | 0 | 0 | 1 | 10276 | /home/jiachen/TRELLIS/experiments/nonorthogonal_gpt_orthoview/outputs/data_survey/vggt_spatial_adapter/support_guidance_a42_component_aware_cleanup/gso_008_input4/mesh_lock_r018_patch030_b024_visible_prior_trellis2_raw.obj |
可视化结果
实验结论
- A43 支持 A42 的拓扑有效性：A42 达到 watertight=100.0%、boundary=0.0、nonmanifold=0.0、components=1.0。
- A43 也暴露 A42 的几何上限：A42 GT Chamfer=0.08827，只比 TRELLIS.2 raw 的 0.09230 好一点，仍明显落后 TRELLIS raw 的 0.05540。
- A40/A41/A42 的 GT F@5 都在约 0.4543 附近，说明 UDF/cleanup 主要修拓扑和复杂度，没有带来足够的真实形状恢复。
- 因此 A42 不能被包装为最终质量提升方法；它更适合作为论文中的 topology prior / post-filter 证据，下一步必须把 signed observable constraint 放到生成期或局部 SDF 中。
- 当前验证只有 2 个 focus case，且是 GT retrospective validation；要发论文还需要扩展到 10/100-case，并补充 20-view render 指标。
下一步想法
- A44：Signed local SDF，把 A42 component prior 变成 signed field 的连通性约束，避免 unsigned shell。
- A45：Generation-time observable-state guidance，在 SLat/decoder 前对局部可见冲突区域施加 VGGT depth residual，而不是 mesh 后处理。
- A46：扩展 GT validation 到 A31 10-case 或 Benchmark100 子集，检查 A42 的 topology/geometry trade-off 是否普遍存在。
- A47：补 20-view render 指标：mask IoU、normal consistency、LPIPS/DreamSim，验证 Chamfer 之外的视觉一致性。