实验定位 小样本正结果 / 需扩大验证
A47 是 A44-A47 链条里的关键正信号：A46 证明 patch-derived sign 大多 inward 且伤 GT，A47 改用输入视角 VGGT depth residual 来决定 active carrier vertex 沿法线 inward/outward。活动区、hard-lock 和拓扑 carrier 都沿用 A45/A46。
结果首次同时保持拓扑并改善 GT：depth_outward_only Chamfer 从 carrier 的 0.08950 降到 0.08823，F@5 从 0.4483 提到 0.4539，同时 watertight=100.0%、components=1。更重要的是，depth-derived outward ratio=45.3%，明显不同于 A46 patch-sign 的约 3.5%。
实验设计（Image-2 风格）
模块设计（Image-2 风格）
Depth-sign Inference
- 把 active carrier vertices 投影到 4 个输入视角。
- carrier raster depth 与 VGGT reference depth 做 robust scale align。
- 通过 depth residual 和 normal-depth derivative 推出 signed normal step。
- 多视角 proposal 取 median，再做 soft/strong/outward/inward ablation。
Interpretation
A47 不再从 replacement patch 最近邻猜方向，而是让输入视角 depth residual 决定内外。这直接回答 A46 的问题：patch sign 不可靠，但 view depth sign 有可能是正确的 observable-state guidance。
实验结果（表格）
聚合结果
| 变体 | GT Chamfer↓ | F@5↑ | 水密率↑ | components↓ | visible p95↓ | conflict p95↓ | patch p95↓ | depth valid active | outward step | p95 move | 结论 |
|---|---|---|---|---|---|---|---|---|---|---|---|
| A42 carrier control | 0.08950 | 0.4483 | 100.0% | 1 | 0.03228 | 0.03759 | 0.03536 | 0.0% | 0.0% | 0.00000 | carrier baseline |
| depth signed soft | 0.08922 | 0.4486 | 100.0% | 1 | 0.03154 | 0.03626 | 0.03823 | 43.6% | 45.3% | 0.01000 | 混合 |
| depth signed strong | 0.08884 | 0.4533 | 100.0% | 1 | 0.03111 | 0.03512 | 0.03731 | 43.6% | 45.3% | 0.01400 | GT 小幅正向 |
| depth outward only | 0.08823 | 0.4539 | 100.0% | 1 | 0.03284 | 0.04074 | 0.04183 | 43.6% | 45.3% | 0.01400 | GT 正向 / proxy 反向 |
| depth inward only | 0.09037 | 0.4452 | 100.0% | 1 | 0.03086 | 0.03392 | 0.03372 | 43.6% | 0.0% | 0.01400 | proxy 好但 GT 差 |
| depth signed smooth | 0.08927 | 0.4501 | 100.0% | 1 | 0.03139 | 0.03639 | 0.03838 | 43.6% | 45.3% | 0.01000 | 混合 |
逐 case 输出
| case | 变体 | GT Chamfer↓ | F@5↑ | watertight | components | conflict p95 | depth valid active | outward step | mean depth residual | OBJ |
|---|---|---|---|---|---|---|---|---|---|---|
| gso_002_input4 | A42 carrier control | 0.04591 | 0.6879 | True | 1 | 0.03551 | 0.0% | 0.0% | - | /home/jiachen/TRELLIS/experiments/nonorthogonal_gpt_orthoview/outputs/data_survey/vggt_spatial_adapter/support_guidance_a47_depth_sign_residual/gso_002_input4/mesh_carrier_control.obj |
| gso_002_input4 | depth signed soft | 0.04603 | 0.6881 | True | 1 | 0.03403 | 18.8% | 56.5% | 0.19666 | /home/jiachen/TRELLIS/experiments/nonorthogonal_gpt_orthoview/outputs/data_survey/vggt_spatial_adapter/support_guidance_a47_depth_sign_residual/gso_002_input4/mesh_depth_signed_soft.obj |
| gso_002_input4 | depth signed strong | 0.04571 | 0.6946 | True | 1 | 0.03396 | 18.8% | 56.5% | 0.19666 | /home/jiachen/TRELLIS/experiments/nonorthogonal_gpt_orthoview/outputs/data_survey/vggt_spatial_adapter/support_guidance_a47_depth_sign_residual/gso_002_input4/mesh_depth_signed_strong.obj |
| gso_002_input4 | depth outward only | 0.04579 | 0.6928 | True | 1 | 0.03849 | 18.8% | 56.5% | 0.19666 | /home/jiachen/TRELLIS/experiments/nonorthogonal_gpt_orthoview/outputs/data_survey/vggt_spatial_adapter/support_guidance_a47_depth_sign_residual/gso_002_input4/mesh_depth_outward_only.obj |
| gso_002_input4 | depth inward only | 0.04576 | 0.6896 | True | 1 | 0.03263 | 18.8% | 0.0% | 0.19666 | /home/jiachen/TRELLIS/experiments/nonorthogonal_gpt_orthoview/outputs/data_survey/vggt_spatial_adapter/support_guidance_a47_depth_sign_residual/gso_002_input4/mesh_depth_inward_only.obj |
| gso_002_input4 | depth signed smooth | 0.04569 | 0.6924 | True | 1 | 0.03401 | 18.8% | 56.5% | 0.19666 | /home/jiachen/TRELLIS/experiments/nonorthogonal_gpt_orthoview/outputs/data_survey/vggt_spatial_adapter/support_guidance_a47_depth_sign_residual/gso_002_input4/mesh_depth_signed_smooth.obj |
| gso_008_input4 | A42 carrier control | 0.13308 | 0.2088 | True | 1 | 0.03966 | 0.0% | 0.0% | - | /home/jiachen/TRELLIS/experiments/nonorthogonal_gpt_orthoview/outputs/data_survey/vggt_spatial_adapter/support_guidance_a47_depth_sign_residual/gso_008_input4/mesh_carrier_control.obj |
| gso_008_input4 | depth signed soft | 0.13240 | 0.2092 | True | 1 | 0.03849 | 68.4% | 34.1% | 0.09194 | /home/jiachen/TRELLIS/experiments/nonorthogonal_gpt_orthoview/outputs/data_survey/vggt_spatial_adapter/support_guidance_a47_depth_sign_residual/gso_008_input4/mesh_depth_signed_soft.obj |
| gso_008_input4 | depth signed strong | 0.13197 | 0.2119 | True | 1 | 0.03629 | 68.4% | 34.1% | 0.09194 | /home/jiachen/TRELLIS/experiments/nonorthogonal_gpt_orthoview/outputs/data_survey/vggt_spatial_adapter/support_guidance_a47_depth_sign_residual/gso_008_input4/mesh_depth_signed_strong.obj |
| gso_008_input4 | depth outward only | 0.13066 | 0.2151 | True | 1 | 0.04298 | 68.4% | 34.1% | 0.09194 | /home/jiachen/TRELLIS/experiments/nonorthogonal_gpt_orthoview/outputs/data_survey/vggt_spatial_adapter/support_guidance_a47_depth_sign_residual/gso_008_input4/mesh_depth_outward_only.obj |
| gso_008_input4 | depth inward only | 0.13498 | 0.2007 | True | 1 | 0.03520 | 68.4% | 0.0% | 0.09194 | /home/jiachen/TRELLIS/experiments/nonorthogonal_gpt_orthoview/outputs/data_survey/vggt_spatial_adapter/support_guidance_a47_depth_sign_residual/gso_008_input4/mesh_depth_inward_only.obj |
| gso_008_input4 | depth signed smooth | 0.13285 | 0.2078 | True | 1 | 0.03878 | 68.4% | 34.1% | 0.09194 | /home/jiachen/TRELLIS/experiments/nonorthogonal_gpt_orthoview/outputs/data_survey/vggt_spatial_adapter/support_guidance_a47_depth_sign_residual/gso_008_input4/mesh_depth_signed_smooth.obj |
可视化结果
实验结论
- A47 是 A44-A47 中第一条几何正向结果：depth_outward_only 同时保持水密/单组件，并降低 GT Chamfer、提高 F@5。
- depth_signed_strong 也小幅正向：Chamfer=0.08884，F@5=0.4533，说明 depth sign 不是只在 outward-only trick 下有效。
- 最反直觉但最重要的发现：depth_inward_only 的 conflict proxy p95=0.03392 最好之一，但 Chamfer=0.09037 最差；这说明 surface proxy 会奖励错误的 inward shrink。
- A47 支持新的论文主张：observable-state guidance 需要 view-depth-derived sign，而不是 patch-nearest 或 surface-proxy-derived sign。
- 限制：当前仍是 2 个 focus case、低分辨率 depth sign、GT retrospective validation；必须扩展到 10/100-case，并把 depth sign 变成无 GT selection gate。
下一步想法
- A48：把 depth_outward_only / depth_signed_strong 扩展到 A31 10-case，验证小样本正结果是否稳定。
- A49：Depth-sign confidence gate，只在多视角 depth proposal 符号一致、normal derivative 稳定、VGGT confidence 足够高时移动。
- A50：Depth sign + local optimization，把 outward/signed proposal 作为 data term，visible hard-lock 作为 Dirichlet boundary，加入 Laplacian regularization。
- A51：补 20-view render metrics，确认 A47 的 Chamfer/F@5 改善是否也转化为 mask/normal/LPIPS 视觉收益。