实验定位 routed depth-sign 混合偏正
A52 是 A47/A48 从 2-case 走向 10-case 的第一次真正路由执行：它不再盲目编辑全部对象，而是先遵守 A51 的 no-GT readiness route。ready 的 gso_000/gso_002/gso_008 进入 depth-sign residual，low-conflict 或 carrier-not-ready 的 7 个 case 保持 carrier。
结果：completed=10/10，active residual=3，defer=7。selected mean Chamfer 从 carrier 的 0.08208 到 0.08199，Δ=-0.000083；全 10-case ΔF@5=0.001326，active case ΔF@5=0.004420。
实验设计（Image-2 风格）
模块设计（Image-2 风格）
Routed Selector
- A51 defer route 优先：低冲突、低置信或坏 topology 不编辑。
- active case 使用 A50 conflict face mask 作为活动区。
- VGGT depth residual 决定法线方向和步长。
- 无 GT selector 在 signed-strong / outward-only / carrier 之间选择；GT 只做事后审计。
Key Difference from A47
A47 证明 depth sign 在 2-case 上有信号；A52 验证的是 pipeline 层面的 reject option。它允许“不动”成为合法输出，防止 A44-A46 里观察到的过度投影和错误内缩。
实验结果（表格）
全 10-case routed selection
| case | A51 route | active | selected | selected reasons | carrier Chamfer | selected Chamfer | Δ Chamfer | carrier F@5 | selected F@5 | Δ F@5 | conflict ratio | selected OBJ |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| gso_000_input4 | ready_standard_depth_sign | True | depth_signed_strong | ['signed_depth_support_available'] | 0.06594 | 0.06517 | -0.000777 | 0.4978 | 0.4981 | 0.000340 | 0.1306 | /home/jiachen/TRELLIS/experiments/nonorthogonal_gpt_orthoview/outputs/data_survey/vggt_spatial_adapter/support_guidance_a52_routed_10case_depth_sign_residual/gso_000_input4/mesh_depth_signed_strong.obj |
| gso_001_input4 | ready_but_defer_low_conflict | False | carrier_control | ['a51_defer_route', 'low_conflict_carrier_only_preferred'] | 0.08978 | 0.08978 | 0.000000 | 0.3288 | 0.3288 | 0.000000 | 0.0290 | /home/jiachen/TRELLIS/experiments/nonorthogonal_gpt_orthoview/outputs/data_survey/vggt_spatial_adapter/support_guidance_a52_routed_10case_depth_sign_residual/gso_001_input4/mesh_carrier_control.obj |
| gso_002_input4 | ready_high_value_depth_sign | True | depth_signed_strong | ['signed_depth_support_available'] | 0.04951 | 0.04820 | -0.001311 | 0.6537 | 0.6606 | 0.006943 | 0.2018 | /home/jiachen/TRELLIS/experiments/nonorthogonal_gpt_orthoview/outputs/data_survey/vggt_spatial_adapter/support_guidance_a52_routed_10case_depth_sign_residual/gso_002_input4/mesh_depth_signed_strong.obj |
| gso_003_input4 | defer_carrier_not_ready | False | carrier_control | ['a51_defer_route', 'low_face_vote_confidence'] | 0.04061 | 0.04061 | 0.000000 | 0.7065 | 0.7065 | 0.000000 | 0.0897 | /home/jiachen/TRELLIS/experiments/nonorthogonal_gpt_orthoview/outputs/data_survey/vggt_spatial_adapter/support_guidance_a52_routed_10case_depth_sign_residual/gso_003_input4/mesh_carrier_control.obj |
| gso_004_input4 | ready_but_defer_low_conflict | False | carrier_control | ['a51_defer_route', 'low_conflict_carrier_only_preferred'] | 0.10814 | 0.10814 | 0.000000 | 0.2178 | 0.2178 | 0.000000 | 0.0038 | /home/jiachen/TRELLIS/experiments/nonorthogonal_gpt_orthoview/outputs/data_survey/vggt_spatial_adapter/support_guidance_a52_routed_10case_depth_sign_residual/gso_004_input4/mesh_carrier_control.obj |
| gso_005_input4 | ready_but_defer_low_conflict | False | carrier_control | ['a51_defer_route', 'low_conflict_carrier_only_preferred'] | 0.06968 | 0.06968 | 0.000000 | 0.3412 | 0.3412 | 0.000000 | 0.0000 | /home/jiachen/TRELLIS/experiments/nonorthogonal_gpt_orthoview/outputs/data_survey/vggt_spatial_adapter/support_guidance_a52_routed_10case_depth_sign_residual/gso_005_input4/mesh_carrier_control.obj |
| gso_006_input4 | defer_carrier_not_ready | False | carrier_control | ['a51_defer_route', 'carrier_topology_not_valid'] | 0.07259 | 0.07259 | 0.000000 | 0.4901 | 0.4901 | 0.000000 | 0.4054 | /home/jiachen/TRELLIS/experiments/nonorthogonal_gpt_orthoview/outputs/data_survey/vggt_spatial_adapter/support_guidance_a52_routed_10case_depth_sign_residual/gso_006_input4/mesh_carrier_control.obj |
| gso_007_input4 | ready_but_defer_low_conflict | False | carrier_control | ['a51_defer_route', 'low_conflict_carrier_only_preferred'] | 0.15702 | 0.15702 | 0.000000 | 0.1648 | 0.1648 | 0.000000 | 0.0264 | /home/jiachen/TRELLIS/experiments/nonorthogonal_gpt_orthoview/outputs/data_survey/vggt_spatial_adapter/support_guidance_a52_routed_10case_depth_sign_residual/gso_007_input4/mesh_carrier_control.obj |
| gso_008_input4 | ready_high_value_depth_sign | True | depth_outward_only | ['high_conflict_high_outward_depth_support'] | 0.13568 | 0.13694 | 0.001259 | 0.1891 | 0.1951 | 0.005978 | 0.7050 | /home/jiachen/TRELLIS/experiments/nonorthogonal_gpt_orthoview/outputs/data_survey/vggt_spatial_adapter/support_guidance_a52_routed_10case_depth_sign_residual/gso_008_input4/mesh_depth_outward_only.obj |
| gso_009_input4 | ready_but_defer_low_conflict | False | carrier_control | ['a51_defer_route', 'low_conflict_carrier_only_preferred'] | 0.03181 | 0.03181 | 0.000000 | 0.7983 | 0.7983 | 0.000000 | 0.0176 | /home/jiachen/TRELLIS/experiments/nonorthogonal_gpt_orthoview/outputs/data_survey/vggt_spatial_adapter/support_guidance_a52_routed_10case_depth_sign_residual/gso_009_input4/mesh_carrier_control.obj |
active residual cases
| case | A51 route | active | selected | selected reasons | carrier Chamfer | selected Chamfer | Δ Chamfer | carrier F@5 | selected F@5 | Δ F@5 | conflict ratio | selected OBJ |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| gso_000_input4 | ready_standard_depth_sign | True | depth_signed_strong | ['signed_depth_support_available'] | 0.06594 | 0.06517 | -0.000777 | 0.4978 | 0.4981 | 0.000340 | 0.1306 | /home/jiachen/TRELLIS/experiments/nonorthogonal_gpt_orthoview/outputs/data_survey/vggt_spatial_adapter/support_guidance_a52_routed_10case_depth_sign_residual/gso_000_input4/mesh_depth_signed_strong.obj |
| gso_002_input4 | ready_high_value_depth_sign | True | depth_signed_strong | ['signed_depth_support_available'] | 0.04951 | 0.04820 | -0.001311 | 0.6537 | 0.6606 | 0.006943 | 0.2018 | /home/jiachen/TRELLIS/experiments/nonorthogonal_gpt_orthoview/outputs/data_survey/vggt_spatial_adapter/support_guidance_a52_routed_10case_depth_sign_residual/gso_002_input4/mesh_depth_signed_strong.obj |
| gso_008_input4 | ready_high_value_depth_sign | True | depth_outward_only | ['high_conflict_high_outward_depth_support'] | 0.13568 | 0.13694 | 0.001259 | 0.1891 | 0.1951 | 0.005978 | 0.7050 | /home/jiachen/TRELLIS/experiments/nonorthogonal_gpt_orthoview/outputs/data_survey/vggt_spatial_adapter/support_guidance_a52_routed_10case_depth_sign_residual/gso_008_input4/mesh_depth_outward_only.obj |
variant aggregate
| variant | cases | Chamfer↓ | F@5↑ | 水密率 | components | conflict p95 | visible p95 | depth valid active | outward step |
|---|---|---|---|---|---|---|---|---|---|
| carrier_control | 10 | 0.08208 | 0.4388 | 100.0% | 21.3 | inf | 0.01405 | 0.0% | 0.0% |
| depth_inward_counterfactual | 3 | 0.08448 | 0.4541 | 100.0% | 1 | 0.01968 | 0.01661 | 47.3% | 0.0% |
| depth_outward_only | 3 | 0.08403 | 0.4486 | 100.0% | 1 | 0.01767 | 0.01601 | 47.3% | 29.5% |
| depth_signed_strong | 3 | 0.08470 | 0.4529 | 100.0% | 1 | 0.02080 | 0.01769 | 47.3% | 29.5% |
可视化结果
实验结论
- A52 路由机制有效：7/10 case 被安全保持为 carrier，避免对 low-conflict 或 carrier-not-ready case 做无意义 residual。
- active 3 case 中，
gso_000和gso_002同时改善 Chamfer 和 F@5；gso_008F@5 改善但 Chamfer 变差，说明 high-conflict 大区域仍需要更强 selector。 - 全 10-case selected Δ Chamfer=-0.000083、ΔF@5=0.001326；active case Δ Chamfer=-0.000276、ΔF@5=0.004420。
- counterfactual inward 在 active case 的 F@5 有时更高，但这是 GT 事后观察，不能直接部署；它提示下一轮不能只用 outward ratio，而要估计 Chamfer/F@5 tradeoff 或多指标 selector。
- A52 支持论文主张：多视角 depth guidance 需要 reject-option 和 mask-aware activity，不应作为全局后处理套到所有对象上。
下一步想法
- A53：为 gso_008 设计 multi-objective selector，同时看 depth support、outward ratio、visible drift、conflict p95 和 topology，不只按 high conflict 选择 outward。
- A54：补 defer counterfactual：对 low-conflict case 轻量运行 residual，证明 A51 defer 是否真的避免劣化。
- A55：引入 signed step calibration，用 depth residual 的尺度置信区间调节 max_step，而不是固定 0.014。
- A56：把 selected mesh 渲染到 20 个测试视角，增加 mask IoU / normal consistency / LPIPS，检查 A52 的 F@5 改善是否转成视觉收益。