实验定位 无 GT 选择器小样本正结果
A48 不再生成新 mesh，而是把 A47 的候选变体转成部署时可用的 no-GT selector。它比较三种思路：carrier 不动、surface proxy 最小、depth-sign confidence gate。GT 只用于 retrospective validation，不参与选择。
结果非常清楚：depth-sign confidence gate 选择 gso_002→depth_signed_strong、gso_008→depth_outward_only，mean Chamfer=0.08819，几乎贴近 GT oracle=0.08818，相对 carrier 改善 -0.00131。相反，surface_proxy_min 两个 case 都选择 inward-only，Chamfer=0.09037，比 carrier=0.08950 更差。
实验设计（Image-2 风格）
模块设计（Image-2 风格）
Gate Rule
- 高 depth-valid active ratio 且 outward proposal 非零：选择
depth_outward_only。 - 低 depth support 但 outward ratio 足够：选择
depth_signed_strong。 - depth support 不足：回退 carrier。
- 不使用 GT Chamfer/F-score 做选择。
Proxy Trap
surface-proxy selector 会选择 conflict p95 最小的 inward-only。A47/A48 证明这会奖励错误内缩，GT 反而变差。因此 A48 的核心不是调参，而是把“观测符号可靠性”变成候选选择标准。
实验结果（表格）
Selector 聚合
| selector | GT Chamfer↓ | Δ Chamfer vs carrier↓ | F@5↑ | Δ F@5↑ | 水密率↑ | components↓ | selected variants | 结论 |
|---|---|---|---|---|---|---|---|---|
| GT oracle upper bound | 0.08818 | -0.00132 | 0.4537 | 0.0054 | 100.0% | 1 | {'gso_002_input4': 'depth_signed_smooth', 'gso_008_input4': 'depth_outward_only'} | 上界 |
| depth-sign confidence gate | 0.08819 | -0.00131 | 0.4548 | 0.0065 | 100.0% | 1 | {'gso_002_input4': 'depth_signed_strong', 'gso_008_input4': 'depth_outward_only'} | 正向 |
| depth outward all | 0.08823 | -0.00127 | 0.4539 | 0.0056 | 100.0% | 1 | {'gso_002_input4': 'depth_outward_only', 'gso_008_input4': 'depth_outward_only'} | 正向 |
| depth signed-strong all | 0.08884 | -0.00066 | 0.4533 | 0.0049 | 100.0% | 1 | {'gso_002_input4': 'depth_signed_strong', 'gso_008_input4': 'depth_signed_strong'} | 正向但保守 |
| conservative depth gate | 0.08884 | -0.00066 | 0.4533 | 0.0049 | 100.0% | 1 | {'gso_002_input4': 'depth_signed_strong', 'gso_008_input4': 'depth_signed_strong'} | 正向但保守 |
| carrier control | 0.08950 | 0.00000 | 0.4483 | 0.0000 | 100.0% | 1 | {'gso_002_input4': 'carrier_control', 'gso_008_input4': 'carrier_control'} | baseline |
| surface proxy min | 0.09037 | 0.00087 | 0.4452 | -0.0032 | 100.0% | 1 | {'gso_002_input4': 'depth_inward_only', 'gso_008_input4': 'depth_inward_only'} | 选错 / 负向 |
逐 case 选择
| selector | case | selected variant | GT Chamfer↓ | Δ Chamfer | F@5↑ | conflict p95 | visible p95 | depth valid active | outward step | OBJ |
|---|---|---|---|---|---|---|---|---|---|---|
| GT oracle upper bound | gso_002_input4 | depth_signed_smooth | 0.04569 | -0.00022 | 0.6924 | 0.03401 | 0.03277 | 18.8% | 56.5% | /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 |
| GT oracle upper bound | gso_008_input4 | depth_outward_only | 0.13066 | -0.00242 | 0.2151 | 0.04298 | 0.03184 | 68.4% | 34.1% | /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 |
| depth-sign confidence gate | gso_002_input4 | depth_signed_strong | 0.04571 | -0.00020 | 0.6946 | 0.03396 | 0.03273 | 18.8% | 56.5% | /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 |
| depth-sign confidence gate | gso_008_input4 | depth_outward_only | 0.13066 | -0.00242 | 0.2151 | 0.04298 | 0.03184 | 68.4% | 34.1% | /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 |
| depth outward all | gso_002_input4 | depth_outward_only | 0.04579 | -0.00012 | 0.6928 | 0.03849 | 0.03385 | 18.8% | 56.5% | /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 |
| depth outward all | gso_008_input4 | depth_outward_only | 0.13066 | -0.00242 | 0.2151 | 0.04298 | 0.03184 | 68.4% | 34.1% | /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 |
| depth signed-strong all | gso_002_input4 | depth_signed_strong | 0.04571 | -0.00020 | 0.6946 | 0.03396 | 0.03273 | 18.8% | 56.5% | /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 |
| depth signed-strong all | gso_008_input4 | depth_signed_strong | 0.13197 | -0.00111 | 0.2119 | 0.03629 | 0.02950 | 68.4% | 34.1% | /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 |
| conservative depth gate | gso_002_input4 | depth_signed_strong | 0.04571 | -0.00020 | 0.6946 | 0.03396 | 0.03273 | 18.8% | 56.5% | /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 |
| conservative depth gate | gso_008_input4 | depth_signed_strong | 0.13197 | -0.00111 | 0.2119 | 0.03629 | 0.02950 | 68.4% | 34.1% | /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 |
| carrier control | gso_002_input4 | carrier_control | 0.04591 | 0.00000 | 0.6879 | 0.03551 | 0.03314 | 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 |
| carrier control | gso_008_input4 | carrier_control | 0.13308 | 0.00000 | 0.2088 | 0.03966 | 0.03142 | 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 |
| surface proxy min | gso_002_input4 | depth_inward_only | 0.04576 | -0.00015 | 0.6896 | 0.03263 | 0.03251 | 18.8% | 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_depth_inward_only.obj |
| surface proxy min | gso_008_input4 | depth_inward_only | 0.13498 | 0.00189 | 0.2007 | 0.03520 | 0.02921 | 68.4% | 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_depth_inward_only.obj |
可视化结果
实验结论
- A48 把 A47 的正向候选变成一个 no-GT selector：depth-sign confidence gate 的 mean Chamfer=0.08819，与 oracle 的差距只有 0.000010。
- surface proxy selector 明确失败：它选择 inward-only，mean Δ Chamfer=0.00087，说明“proxy 最小”不是可靠目标。
- depth_outward_all 也正向，但 confidence gate 更接近 oracle，因为 gso_002 低 depth support 时选择 signed-strong 更稳。
- 这给论文方法提供一个可部署模块：不是只提出 depth-sign residual，还能用 no-GT confidence rule 决定何时 outward、何时 signed、何时 carrier。
- 限制：仍只有 2 个 carrier/mask case；必须扩展 A34/A35/A42 到 10-case 后重测 selector 的稳定性。
下一步想法
- A49：先把 A34/A35/A42 carrier artifact 扩展到 A31 的 10-case，再跑 A47/A48。
- A50：加入 confidence calibration：depth-valid ratio、outward ratio、multi-view sign consistency、VGGT confidence 分别做 ablation。
- A51：让 selector 输出 defer，而不是必须选一个变体；低置信时保持 carrier。
- A52：把 A48 写成论文中的 candidate router：Observable-State Depth Sign Router。