SceneBenchmark 真实场景评测

同学 week/25 的 VLM critic（run_vlm_critic.py + vlm_critic_ours.yaml）× 我们的 API 端点（avemujica gpt-5.5） · 2026-06-17 · Blender 4.2.9 LTS 带纹理

✔ 已修复：本页整房 overview 此前出现扁平件（地毯/壁架）被拉伸“立起”成竖板的转换器 bug。根因=场景 local_bounds 存的是资产 Y-up 帧（薄轴=Y），而渲染器把 glTF 导入后已转成 Z-up（薄轴=Z）的 mesh 按 local_bounds 逐轴缩放，导致薄轴被放大 ~95×。 修复（场景数据级）：对大扁平件把 local_bounds/size 的薄轴换到 Z（scene_fixed.json，不改同学渲染器源码、不改原 scene.json）。下方 overview 已为修复后干净版——地毯平铺在地面、整房布局正常、家具未被带歪。

一句话：在已按官方 converter 转好的真实卧室场景 hssd_bedroom_plus_articraft（22 件家具）上，用同学最新主线的 critic 真实跑通功能性评测（交互净空 / 可达性 / 功能依赖），VLM 走我们约定的 gpt-5.5。本页另渲了相机拉远的整房 overview（房外上方俯视、隐去天花与近侧墙），确保整个房间布局可见。诚实：该 critic 是 规则优先 + VLM 兜底——几何/布局规则判得了就判，判不了才交 VLM。

整房 overview（相机拉远 · 三角度）

overview NE — 东北俯视：床（深色寝具）+ 床尾长凳 + 衣柜 + 床头柜 + 墙架。

overview SW — 西南俯视：地毯 + 床 + 双床头柜 + 墙架 + 顶灯/烟感小件。

overview SE — 东南俯视：整体布局，衣柜/家电、长凳、墙架、镜子一并入框。

相机置于房间外上方（壳对角外 ~0.55×、高 ~1.7×壳高，fov 52°，target=房间中心），渲染时隐去天花板与近侧两面墙以露出室内布局。带纹理（HSSD PBR：木纹架/长凳腿、深色镜框、黑色寝具、白家具）。

评测汇总 · `hssd_bedroom_plus_articraft`（22 件 · 卧室）

pass

degraded

fail

checks 总计

VLM 真裁决

规则裁决

check 构成：交互净空 ×6 · 可达性 ×5 · 功能依赖 ×7。报告模式：实际 VLM 评审（rule-first 混合）。

逐 check 卡片（18 个）

每张卡 = 该 check 的局部渲染图 + 裁决（pass/degraded/fail）+ 裁决来源（VLM/规则）+ confidence + reason 截断。

interaction_clearance__bedroom_nightstand_0 — passVLM 裁决conf 0.88
`交互净空 · interaction_clearance · nightstand_0`
In the local context render, the nightstand’s top surface is visible and mostly unobstructed for placing or reaching items. No direct blocker is sitting on the usable top face, and the nearby bed frame stays outside the tabletop use zone in the local view.

interaction_clearance__bedroom_nightstand_1 — passVLM 裁决conf 0.82
`交互净空 · interaction_clearance · nightstand_1`
In the local_context render, the highlighted nightstand is visible beside the bed with its top surface exposed; the nearby bed frame and wall do not cover the usable top. In the overview renders, small bedside items are near/above the nightstand area, but…

interaction_clearance__bedroom_wardrobe_0 — degraded规则裁决conf 0.80
`交互净空 · interaction_clearance · wardrobe_0`
Rule clearance: limiting profile `default` has best `front` interaction zone for `bedroom_wardrobe_0` with 0.38 immediate free-space overlap. Nearby/intersecting obstacles: appliance_articraft_0.

blocking：`appliance_articraft_0`

interaction_clearance__bedroom_jewelry_catchall_dish_0__to__bedroom_nightstand_1__bedroom_nightstand_0 — pass规则裁决conf 0.90
`交互净空 · interaction_clearance · jewelry_catchall_dish_0 → nightstand_1 · nightstand_0`
Rule clearance: limiting profile `default` has best `front` interaction zone for `bedroom_jewelry_catchall_dish_0` with 1.00 immediate free-space overlap.

interaction_clearance__bedroom_bed_0__to__bedroom_rug_0__bedroom_nightstand_0 — pass规则裁决conf 0.90
`交互净空 · interaction_clearance · bed_0 → rug_0 · nightstand_0`
Rule clearance: limiting profile `default` has best `front` interaction zone for `bedroom_bed_0` with 1.00 immediate free-space overlap.

interaction_clearance__bedroom_bedroom_bench_0__to__bedroom_shelf_0 — pass规则裁决conf 0.90
`交互净空 · interaction_clearance · bench_0 → shelf_0`
Rule clearance: limiting profile `default` has best `front` interaction zone for `bedroom_bedroom_bench_0` with 1.00 immediate free-space overlap.

spatial_accessibility__bedroom_wardrobe_0 — degraded规则裁决conf 0.82
`可达性 · spatial_accessibility · wardrobe_0`
Rule accessibility: limiting profile `default` has best `front` access zone for `bedroom_wardrobe_0` with 0.38 connected stance overlap and 0.63m minimum reach distance. Minimum reach uses crouch/lean posture. Nearby/intersecting obstacles:…

blocking：`appliance_articraft_0`

spatial_accessibility__bedroom_bed_0__to__bedroom_rug_0__bedroom_nightstand_0 — pass规则裁决conf 0.92
`可达性 · spatial_accessibility · bed_0 → rug_0 · nightstand_0`
Rule accessibility: limiting profile `default` has best `front` access zone for `bedroom_bed_0` with 1.00 connected stance overlap and 0.04m minimum reach distance.

spatial_accessibility__bedroom_bedroom_bench_0__to__bedroom_shelf_0 — pass规则裁决conf 0.92
`可达性 · spatial_accessibility · bench_0 → shelf_0`
Rule accessibility: limiting profile `default` has best `front` access zone for `bedroom_bedroom_bench_0` with 1.00 connected stance overlap and 0.04m minimum reach distance.

spatial_accessibility__bedroom_nightstand_0 — pass规则裁决conf 0.92
`可达性 · spatial_accessibility · nightstand_0`
Rule accessibility: limiting profile `default` has best `front` access zone for `bedroom_nightstand_0` with 0.92 connected stance overlap and 0.43m minimum reach distance. Minimum reach uses crouch/lean posture.

spatial_accessibility__bedroom_nightstand_1 — pass规则裁决conf 0.92
`可达性 · spatial_accessibility · nightstand_1`
Rule accessibility: limiting profile `default` has best `front` access zone for `bedroom_nightstand_1` with 1.00 connected stance overlap and 0.43m minimum reach distance. Minimum reach uses crouch/lean posture.

functional_dependency__bedroom_clear_glass_tumbler_0__to__bedroom_water_carafe_0 — degraded规则裁决conf 0.66
`功能依赖 · functional_dependency · clear_glass_tumbler_0 → water_carafe_0`
Rule dependency `object_on_support`: subject `bedroom_clear_glass_tumbler_0`; selected `bedroom_water_carafe_0`; unified support score selected bbox top fallback `bbox_top`: overlap 0.08, height delta 0.08m.

functional_dependency__bedroom_water_carafe_0__to__bedroom_nightstand_1 — degraded规则裁决conf 0.70
`功能依赖 · functional_dependency · water_carafe_0 → nightstand_1`
Rule dependency `object_on_support`: subject `bedroom_water_carafe_0`; selected `bedroom_nightstand_1`; top support is approximate via bbox fallback `support_region_0`: region overlap 0.54, bbox overlap 0.83, height delta 0.27m.

functional_dependency__bedroom_bed_0__to__bedroom_nightstand_1 — pass规则裁决conf 0.88
`功能依赖 · functional_dependency · bed_0 → nightstand_1`
Rule dependency `seating_to_work_surface`: subject `bedroom_bed_0`; selected `bedroom_nightstand_1`; seat is adjacent to a side surface with gap 0.26m.

functional_dependency__bedroom_alarm_clock_0__to__bedroom_nightstand_0 — pass规则裁决conf 0.82
`功能依赖 · functional_dependency · alarm_clock_0 → nightstand_0`
Rule dependency `object_on_support`: subject `bedroom_alarm_clock_0`; selected `bedroom_nightstand_0`; support region is conservative for a small bedside object, but strong bbox fallback top/edge support is plausible `support_region_0`: region overlap 1.00,…

functional_dependency__bedroom_hardcover_book_0__to__bedroom_nightstand_1 — pass规则裁决conf 0.84
`功能依赖 · functional_dependency · hardcover_book_0 → nightstand_1`
Rule dependency `object_on_support`: subject `bedroom_hardcover_book_0`; selected `bedroom_nightstand_1`; support region is conservative, but bbox top/edge support is plausible: region overlap 0.00, bbox overlap 1.00, height delta 0.07m.

functional_dependency__bedroom_jewelry_catchall_dish_0__to__bedroom_nightstand_1 — pass规则裁决conf 0.82
`功能依赖 · functional_dependency · jewelry_catchall_dish_0 → nightstand_1`
Rule dependency `object_on_support`: subject `bedroom_jewelry_catchall_dish_0`; selected `bedroom_nightstand_1`; support region is conservative for a small bedside object, but strong bbox fallback top/edge support is plausible `support_region_0`: region…

functional_dependency__bedroom_paperback_book_0__to__bedroom_nightstand_0 — pass规则裁决conf 0.90
`功能依赖 · functional_dependency · paperback_book_0 → nightstand_0`
Rule dependency `object_on_support`: subject `bedroom_paperback_book_0`; selected `bedroom_nightstand_0`; matched top_surface `support_region_0` with overlap 0.96, height delta 0.07m, and clearance 1.00m.

VLM 真裁决示例

“In the local context render, the nightstand’s top surface is visible and mostly unobstructed for placing or reaching items. No direct blocker is sitting on the usable top face…”
— gpt-5.5 对 nightstand_0 顶面占用的视觉裁决（引用了实际渲染图，证明 VLM 真被调用，conf 0.88）

诚实说明 & 坑

critic 为 rule-first 混合：几何/布局规则能判的不调 VLM；本场景 16/18 规则裁决、2/18 VLM 真裁决（均为床头柜顶面占用这类视觉歧义 check）。
替换严格按官方：直接用官方 converter 产出的 scene.json，未注入我们的实验件。
拉远 overview 渲染坑：默认房间壳是闭合盒（带天花板），房外俯视只能看到盒子外壳；解法=渲染时把 ceiling 与朝向相机的两面墙 hide_render，露出室内。已逐张亲眼确认整房布局可见。
VLM 端点坑：环境 ALL_PROXY=socks5h 会让 httpx 报 socks 错 → 须 unset 全部代理直连；config 用 avemujica.moe/v1 + AVEMUJICA_API_KEY。
渲染须用 Blender 4.2.9 LTS（系统 3.0.1 的 EEVEE 不渲 glTF 纹理）。