SceneTransporter · experiment detail
2026-04-19 22:30:00 UTC

SceneTransporter Structure Probe on PartNeXt

严格按 SceneTransporter 的核心思想做了一个最小 probe:不直接碰 3D 生成器,而是只比较 patch-to-part assignment。本次在 6 个 PartNeXt rendered inputs 上对比了 `SAM`、`DINO patch KMeans`、`cosine routing`、`OT-noedge` 和 `SceneTransporter 风格 OT+edge`。结果并不自动站在 OT 这一边:均值上 `KMeans/cosine = 0.760`,`OT+edge = 0.708`,`OT-noedge = 0.702`,`SAM = 0.619`。这说明论文里的 assignment 约束思想很强,但在我们当前“无 compositional latent、只有图像 patch 特征”的简化设定里,还不能直接复制出它的优势。

SceneTransporterPartNeXtstructure-probeOTmask-assignment
2026-04-19 22:30:00 UTCTimestamp
6Assets
activeStatus
SceneTransporter Structure Probe on PartNeXt cover image
Assets
Interactive Asset

Knife_00602ef508784e5384665aacaaf1f3a0

GT 顶层 part 数为 2,名称是 `Blade / Handle`。最好的结构分离方法是 `kmeans`,matched IoU = 1.000。为了避免原始 patch mask 看起来像大方块,页面里额外放了 `boundary overlay` 和 `watershed-refined overlay`。完整对比分数:kmeans=1.000 / cosine=1.000 / ot_noedge=1.000 / ot_edge=0.945 / sam=0.498。

2GT parts
kmeansbest method
1.000best matched IoU
Knife_00602ef508784e5384665aacaaf1f3a0 GT / SAM / KMeans / cosine / OT-noedge / OT-edge
GT / SAM / KMeans / cosine / OT-noedge / OT-edge
Knife_00602ef508784e5384665aacaaf1f3a0 GT mask overlay
GT mask overlay
Knife_00602ef508784e5384665aacaaf1f3a0 sam overlay (blocky patch view, IoU=0.498)
sam overlay (blocky patch view, IoU=0.498)
Knife_00602ef508784e5384665aacaaf1f3a0 sam boundary-only overlay
sam boundary-only overlay
Knife_00602ef508784e5384665aacaaf1f3a0 sam watershed-refined overlay
sam watershed-refined overlay
Knife_00602ef508784e5384665aacaaf1f3a0 kmeans overlay (blocky patch view, IoU=1.000)
kmeans overlay (blocky patch view, IoU=1.000)
Knife_00602ef508784e5384665aacaaf1f3a0 kmeans boundary-only overlay
kmeans boundary-only overlay
Knife_00602ef508784e5384665aacaaf1f3a0 kmeans watershed-refined overlay
kmeans watershed-refined overlay
Knife_00602ef508784e5384665aacaaf1f3a0 cosine overlay (blocky patch view, IoU=1.000)
cosine overlay (blocky patch view, IoU=1.000)
Knife_00602ef508784e5384665aacaaf1f3a0 cosine boundary-only overlay
cosine boundary-only overlay
Knife_00602ef508784e5384665aacaaf1f3a0 cosine watershed-refined overlay
cosine watershed-refined overlay
Knife_00602ef508784e5384665aacaaf1f3a0 ot_noedge overlay (blocky patch view, IoU=1.000)
ot_noedge overlay (blocky patch view, IoU=1.000)
Knife_00602ef508784e5384665aacaaf1f3a0 ot_noedge boundary-only overlay
ot_noedge boundary-only overlay
Knife_00602ef508784e5384665aacaaf1f3a0 ot_noedge watershed-refined overlay
ot_noedge watershed-refined overlay
Knife_00602ef508784e5384665aacaaf1f3a0 ot_edge overlay (blocky patch view, IoU=0.945)
ot_edge overlay (blocky patch view, IoU=0.945)
Knife_00602ef508784e5384665aacaaf1f3a0 ot_edge boundary-only overlay
ot_edge boundary-only overlay
Knife_00602ef508784e5384665aacaaf1f3a0 ot_edge watershed-refined overlay
ot_edge watershed-refined overlay
Knife_00602ef508784e5384665aacaaf1f3a0 context 3D
右侧 3D 是同一样例的上下文资产;左侧新的 boundary / refined 视图更适合看结构边界,而不是看 patch 马赛克。
Interactive Asset

Toilet_01b31c7fb7bd41ac8019ffc994b22b60

GT 顶层 part 数为 3,名称是 `Tank / Toilet Lid / Toilet Base`。最好的结构分离方法是 `sam`,matched IoU = 0.344。为了避免原始 patch mask 看起来像大方块,页面里额外放了 `boundary overlay` 和 `watershed-refined overlay`。完整对比分数:sam=0.344 / kmeans=0.325 / cosine=0.325 / ot_noedge=0.325 / ot_edge=0.299。

3GT parts
sambest method
0.344best matched IoU
Toilet_01b31c7fb7bd41ac8019ffc994b22b60 GT / SAM / KMeans / cosine / OT-noedge / OT-edge
GT / SAM / KMeans / cosine / OT-noedge / OT-edge
Toilet_01b31c7fb7bd41ac8019ffc994b22b60 GT mask overlay
GT mask overlay
Toilet_01b31c7fb7bd41ac8019ffc994b22b60 sam overlay (blocky patch view, IoU=0.344)
sam overlay (blocky patch view, IoU=0.344)
Toilet_01b31c7fb7bd41ac8019ffc994b22b60 sam boundary-only overlay
sam boundary-only overlay
Toilet_01b31c7fb7bd41ac8019ffc994b22b60 sam watershed-refined overlay
sam watershed-refined overlay
Toilet_01b31c7fb7bd41ac8019ffc994b22b60 kmeans overlay (blocky patch view, IoU=0.325)
kmeans overlay (blocky patch view, IoU=0.325)
Toilet_01b31c7fb7bd41ac8019ffc994b22b60 kmeans boundary-only overlay
kmeans boundary-only overlay
Toilet_01b31c7fb7bd41ac8019ffc994b22b60 kmeans watershed-refined overlay
kmeans watershed-refined overlay
Toilet_01b31c7fb7bd41ac8019ffc994b22b60 cosine overlay (blocky patch view, IoU=0.325)
cosine overlay (blocky patch view, IoU=0.325)
Toilet_01b31c7fb7bd41ac8019ffc994b22b60 cosine boundary-only overlay
cosine boundary-only overlay
Toilet_01b31c7fb7bd41ac8019ffc994b22b60 cosine watershed-refined overlay
cosine watershed-refined overlay
Toilet_01b31c7fb7bd41ac8019ffc994b22b60 ot_noedge overlay (blocky patch view, IoU=0.325)
ot_noedge overlay (blocky patch view, IoU=0.325)
Toilet_01b31c7fb7bd41ac8019ffc994b22b60 ot_noedge boundary-only overlay
ot_noedge boundary-only overlay
Toilet_01b31c7fb7bd41ac8019ffc994b22b60 ot_noedge watershed-refined overlay
ot_noedge watershed-refined overlay
Toilet_01b31c7fb7bd41ac8019ffc994b22b60 ot_edge overlay (blocky patch view, IoU=0.299)
ot_edge overlay (blocky patch view, IoU=0.299)
Toilet_01b31c7fb7bd41ac8019ffc994b22b60 ot_edge boundary-only overlay
ot_edge boundary-only overlay
Toilet_01b31c7fb7bd41ac8019ffc994b22b60 ot_edge watershed-refined overlay
ot_edge watershed-refined overlay
Toilet_01b31c7fb7bd41ac8019ffc994b22b60 context 3D
右侧 3D 是同一样例的上下文资产;左侧新的 boundary / refined 视图更适合看结构边界,而不是看 patch 马赛克。
Interactive Asset

Monitor_01ce620e70ff40708eb4a1b04f4a828e

GT 顶层 part 数为 2,名称是 `Display / Stand`。最好的结构分离方法是 `sam`,matched IoU = 1.000。为了避免原始 patch mask 看起来像大方块,页面里额外放了 `boundary overlay` 和 `watershed-refined overlay`。完整对比分数:sam=1.000 / ot_edge=0.954 / kmeans=0.889 / cosine=0.889 / ot_noedge=0.889。

2GT parts
sambest method
1.000best matched IoU
Monitor_01ce620e70ff40708eb4a1b04f4a828e GT / SAM / KMeans / cosine / OT-noedge / OT-edge
GT / SAM / KMeans / cosine / OT-noedge / OT-edge
Monitor_01ce620e70ff40708eb4a1b04f4a828e GT mask overlay
GT mask overlay
Monitor_01ce620e70ff40708eb4a1b04f4a828e sam overlay (blocky patch view, IoU=1.000)
sam overlay (blocky patch view, IoU=1.000)
Monitor_01ce620e70ff40708eb4a1b04f4a828e sam boundary-only overlay
sam boundary-only overlay
Monitor_01ce620e70ff40708eb4a1b04f4a828e sam watershed-refined overlay
sam watershed-refined overlay
Monitor_01ce620e70ff40708eb4a1b04f4a828e kmeans overlay (blocky patch view, IoU=0.889)
kmeans overlay (blocky patch view, IoU=0.889)
Monitor_01ce620e70ff40708eb4a1b04f4a828e kmeans boundary-only overlay
kmeans boundary-only overlay
Monitor_01ce620e70ff40708eb4a1b04f4a828e kmeans watershed-refined overlay
kmeans watershed-refined overlay
Monitor_01ce620e70ff40708eb4a1b04f4a828e cosine overlay (blocky patch view, IoU=0.889)
cosine overlay (blocky patch view, IoU=0.889)
Monitor_01ce620e70ff40708eb4a1b04f4a828e cosine boundary-only overlay
cosine boundary-only overlay
Monitor_01ce620e70ff40708eb4a1b04f4a828e cosine watershed-refined overlay
cosine watershed-refined overlay
Monitor_01ce620e70ff40708eb4a1b04f4a828e ot_noedge overlay (blocky patch view, IoU=0.889)
ot_noedge overlay (blocky patch view, IoU=0.889)
Monitor_01ce620e70ff40708eb4a1b04f4a828e ot_noedge boundary-only overlay
ot_noedge boundary-only overlay
Monitor_01ce620e70ff40708eb4a1b04f4a828e ot_noedge watershed-refined overlay
ot_noedge watershed-refined overlay
Monitor_01ce620e70ff40708eb4a1b04f4a828e ot_edge overlay (blocky patch view, IoU=0.954)
ot_edge overlay (blocky patch view, IoU=0.954)
Monitor_01ce620e70ff40708eb4a1b04f4a828e ot_edge boundary-only overlay
ot_edge boundary-only overlay
Monitor_01ce620e70ff40708eb4a1b04f4a828e ot_edge watershed-refined overlay
ot_edge watershed-refined overlay
Monitor_01ce620e70ff40708eb4a1b04f4a828e context 3D
右侧 3D 是同一样例的上下文资产;左侧新的 boundary / refined 视图更适合看结构边界,而不是看 patch 马赛克。
Interactive Asset

Guitar_553a53ba86804d4da6e51946a6011b0e

GT 顶层 part 数为 2,名称是 `String Components / Guitar Main Components`。最好的结构分离方法是 `sam`,matched IoU = 1.000。为了避免原始 patch mask 看起来像大方块,页面里额外放了 `boundary overlay` 和 `watershed-refined overlay`。完整对比分数:sam=1.000 / kmeans=0.643 / cosine=0.643 / ot_noedge=0.643 / ot_edge=0.643。

2GT parts
sambest method
1.000best matched IoU
Guitar_553a53ba86804d4da6e51946a6011b0e GT / SAM / KMeans / cosine / OT-noedge / OT-edge
GT / SAM / KMeans / cosine / OT-noedge / OT-edge
Guitar_553a53ba86804d4da6e51946a6011b0e GT mask overlay
GT mask overlay
Guitar_553a53ba86804d4da6e51946a6011b0e sam overlay (blocky patch view, IoU=1.000)
sam overlay (blocky patch view, IoU=1.000)
Guitar_553a53ba86804d4da6e51946a6011b0e sam boundary-only overlay
sam boundary-only overlay
Guitar_553a53ba86804d4da6e51946a6011b0e sam watershed-refined overlay
sam watershed-refined overlay
Guitar_553a53ba86804d4da6e51946a6011b0e kmeans overlay (blocky patch view, IoU=0.643)
kmeans overlay (blocky patch view, IoU=0.643)
Guitar_553a53ba86804d4da6e51946a6011b0e kmeans boundary-only overlay
kmeans boundary-only overlay
Guitar_553a53ba86804d4da6e51946a6011b0e kmeans watershed-refined overlay
kmeans watershed-refined overlay
Guitar_553a53ba86804d4da6e51946a6011b0e cosine overlay (blocky patch view, IoU=0.643)
cosine overlay (blocky patch view, IoU=0.643)
Guitar_553a53ba86804d4da6e51946a6011b0e cosine boundary-only overlay
cosine boundary-only overlay
Guitar_553a53ba86804d4da6e51946a6011b0e cosine watershed-refined overlay
cosine watershed-refined overlay
Guitar_553a53ba86804d4da6e51946a6011b0e ot_noedge overlay (blocky patch view, IoU=0.643)
ot_noedge overlay (blocky patch view, IoU=0.643)
Guitar_553a53ba86804d4da6e51946a6011b0e ot_noedge boundary-only overlay
ot_noedge boundary-only overlay
Guitar_553a53ba86804d4da6e51946a6011b0e ot_noedge watershed-refined overlay
ot_noedge watershed-refined overlay
Guitar_553a53ba86804d4da6e51946a6011b0e ot_edge overlay (blocky patch view, IoU=0.643)
ot_edge overlay (blocky patch view, IoU=0.643)
Guitar_553a53ba86804d4da6e51946a6011b0e ot_edge boundary-only overlay
ot_edge boundary-only overlay
Guitar_553a53ba86804d4da6e51946a6011b0e ot_edge watershed-refined overlay
ot_edge watershed-refined overlay
Guitar_553a53ba86804d4da6e51946a6011b0e context 3D
右侧 3D 是同一样例的上下文资产;左侧新的 boundary / refined 视图更适合看结构边界,而不是看 patch 马赛克。
Interactive Asset

Teapot_7641732252ad47a5af0828d4f471338b

GT 顶层 part 数为 4,名称是 `Body / Spout / Lid / Handle`。最好的结构分离方法是 `kmeans`,matched IoU = 0.795。为了避免原始 patch mask 看起来像大方块,页面里额外放了 `boundary overlay` 和 `watershed-refined overlay`。完整对比分数:kmeans=0.795 / cosine=0.795 / ot_edge=0.586 / ot_noedge=0.538 / sam=0.369。

4GT parts
kmeansbest method
0.795best matched IoU
Teapot_7641732252ad47a5af0828d4f471338b GT / SAM / KMeans / cosine / OT-noedge / OT-edge
GT / SAM / KMeans / cosine / OT-noedge / OT-edge
Teapot_7641732252ad47a5af0828d4f471338b GT mask overlay
GT mask overlay
Teapot_7641732252ad47a5af0828d4f471338b sam overlay (blocky patch view, IoU=0.369)
sam overlay (blocky patch view, IoU=0.369)
Teapot_7641732252ad47a5af0828d4f471338b sam boundary-only overlay
sam boundary-only overlay
Teapot_7641732252ad47a5af0828d4f471338b sam watershed-refined overlay
sam watershed-refined overlay
Teapot_7641732252ad47a5af0828d4f471338b kmeans overlay (blocky patch view, IoU=0.795)
kmeans overlay (blocky patch view, IoU=0.795)
Teapot_7641732252ad47a5af0828d4f471338b kmeans boundary-only overlay
kmeans boundary-only overlay
Teapot_7641732252ad47a5af0828d4f471338b kmeans watershed-refined overlay
kmeans watershed-refined overlay
Teapot_7641732252ad47a5af0828d4f471338b cosine overlay (blocky patch view, IoU=0.795)
cosine overlay (blocky patch view, IoU=0.795)
Teapot_7641732252ad47a5af0828d4f471338b cosine boundary-only overlay
cosine boundary-only overlay
Teapot_7641732252ad47a5af0828d4f471338b cosine watershed-refined overlay
cosine watershed-refined overlay
Teapot_7641732252ad47a5af0828d4f471338b ot_noedge overlay (blocky patch view, IoU=0.538)
ot_noedge overlay (blocky patch view, IoU=0.538)
Teapot_7641732252ad47a5af0828d4f471338b ot_noedge boundary-only overlay
ot_noedge boundary-only overlay
Teapot_7641732252ad47a5af0828d4f471338b ot_noedge watershed-refined overlay
ot_noedge watershed-refined overlay
Teapot_7641732252ad47a5af0828d4f471338b ot_edge overlay (blocky patch view, IoU=0.586)
ot_edge overlay (blocky patch view, IoU=0.586)
Teapot_7641732252ad47a5af0828d4f471338b ot_edge boundary-only overlay
ot_edge boundary-only overlay
Teapot_7641732252ad47a5af0828d4f471338b ot_edge watershed-refined overlay
ot_edge watershed-refined overlay
Teapot_7641732252ad47a5af0828d4f471338b context 3D
右侧 3D 是同一样例的上下文资产;左侧新的 boundary / refined 视图更适合看结构边界,而不是看 patch 马赛克。
Interactive Asset

Laptop_Computer_a14d471ffda04d38a1910b9ef87e8dff

GT 顶层 part 数为 3,名称是 `Screen Side / Bottom Side / Hinge`。最好的结构分离方法是 `kmeans`,matched IoU = 0.909。为了避免原始 patch mask 看起来像大方块,页面里额外放了 `boundary overlay` 和 `watershed-refined overlay`。完整对比分数:kmeans=0.909 / cosine=0.909 / ot_noedge=0.818 / ot_edge=0.818 / sam=0.500。

3GT parts
kmeansbest method
0.909best matched IoU
Laptop_Computer_a14d471ffda04d38a1910b9ef87e8dff GT / SAM / KMeans / cosine / OT-noedge / OT-edge
GT / SAM / KMeans / cosine / OT-noedge / OT-edge
Laptop_Computer_a14d471ffda04d38a1910b9ef87e8dff GT mask overlay
GT mask overlay
Laptop_Computer_a14d471ffda04d38a1910b9ef87e8dff sam overlay (blocky patch view, IoU=0.500)
sam overlay (blocky patch view, IoU=0.500)
Laptop_Computer_a14d471ffda04d38a1910b9ef87e8dff sam boundary-only overlay
sam boundary-only overlay
Laptop_Computer_a14d471ffda04d38a1910b9ef87e8dff sam watershed-refined overlay
sam watershed-refined overlay
Laptop_Computer_a14d471ffda04d38a1910b9ef87e8dff kmeans overlay (blocky patch view, IoU=0.909)
kmeans overlay (blocky patch view, IoU=0.909)
Laptop_Computer_a14d471ffda04d38a1910b9ef87e8dff kmeans boundary-only overlay
kmeans boundary-only overlay
Laptop_Computer_a14d471ffda04d38a1910b9ef87e8dff kmeans watershed-refined overlay
kmeans watershed-refined overlay
Laptop_Computer_a14d471ffda04d38a1910b9ef87e8dff cosine overlay (blocky patch view, IoU=0.909)
cosine overlay (blocky patch view, IoU=0.909)
Laptop_Computer_a14d471ffda04d38a1910b9ef87e8dff cosine boundary-only overlay
cosine boundary-only overlay
Laptop_Computer_a14d471ffda04d38a1910b9ef87e8dff cosine watershed-refined overlay
cosine watershed-refined overlay
Laptop_Computer_a14d471ffda04d38a1910b9ef87e8dff ot_noedge overlay (blocky patch view, IoU=0.818)
ot_noedge overlay (blocky patch view, IoU=0.818)
Laptop_Computer_a14d471ffda04d38a1910b9ef87e8dff ot_noedge boundary-only overlay
ot_noedge boundary-only overlay
Laptop_Computer_a14d471ffda04d38a1910b9ef87e8dff ot_noedge watershed-refined overlay
ot_noedge watershed-refined overlay
Laptop_Computer_a14d471ffda04d38a1910b9ef87e8dff ot_edge overlay (blocky patch view, IoU=0.818)
ot_edge overlay (blocky patch view, IoU=0.818)
Laptop_Computer_a14d471ffda04d38a1910b9ef87e8dff ot_edge boundary-only overlay
ot_edge boundary-only overlay
Laptop_Computer_a14d471ffda04d38a1910b9ef87e8dff ot_edge watershed-refined overlay
ot_edge watershed-refined overlay
Laptop_Computer_a14d471ffda04d38a1910b9ef87e8dff context 3D
右侧 3D 是同一样例的上下文资产;左侧新的 boundary / refined 视图更适合看结构边界,而不是看 patch 马赛克。