# EXP-3D-RGBD-NONLEARNED-FUSION-V1

不经学习的 RGB-D 融合基线：上下文深度直接体素化进同一个 16³ 网格、同一 bounds，用同一个固定 renderer 读出；只有 4 个标量在 TRAIN24 上拟合，选中 {'k': 1.0, 'a': 16.0, 'b': 4.0, 'c': -2.0}。所有状态在读取 query GT 之前封存。

## DEV

平均 AbsRel：NLF 0.5958，V5 C1（学习，测得深度）0.5030，V5 C0（学习，只用 RGB）0.6174，常数 0.5070。

- LEARNED_MINUS_NLF = +0.0928，95% CI [+0.0061, +0.1816]，ABOVE
- NLF_MINUS_CONSTANT = -0.0889，95% CI [-0.1991, +0.0049]，NOT_DISTINGUISHABLE
- RGB_ONLY_MINUS_NLF = -0.0216，95% CI [-0.1146, +0.0705]，NOT_DISTINGUISHABLE

## DEV_hit_scenes

平均 AbsRel：NLF 0.5381，V5 C1（学习，测得深度）0.4321，V5 C0（学习，只用 RGB）0.5628，常数 0.4930。

- LEARNED_MINUS_NLF = +0.1061，95% CI [+0.0047, +0.1995]，NOT_DISTINGUISHABLE
- NLF_MINUS_CONSTANT = -0.0451，95% CI [-0.1192, +0.0268]，NOT_DISTINGUISHABLE
- RGB_ONLY_MINUS_NLF = -0.0247，95% CI [-0.1338, +0.0784]，NOT_DISTINGUISHABLE

## FRESH_V1

平均 AbsRel：NLF 0.5278，V5 C1（学习，测得深度）0.4724，V5 C0（学习，只用 RGB）0.4788，常数 0.4935。

- LEARNED_MINUS_NLF = +0.0554，95% CI [-0.0431, +0.1528]，NOT_DISTINGUISHABLE
- NLF_MINUS_CONSTANT = -0.0343，95% CI [-0.2050, +0.1029]，NOT_DISTINGUISHABLE
- RGB_ONLY_MINUS_NLF = +0.0491，95% CI [+0.0277, +0.0829]，ABOVE