We investigate surface modeling of multi-transparent interfaces to achieve accurate light transmission. The interplay of errors in geometric structure, light paths, and observation results during the reconstruction of transparent media with a high number of layers makes this a challenging problem. Some existing techniques address the reconstruction of transparent scenes by reducing the number of layers or simplifying light paths; however, this limits their direct applicability to high-order transparent scenes. We propose GaussTrans, a compact framework for light transport in high-order transparent scenes, consisting of three components: Gaussian-Field Hybrid Surfaces, Transport-Coupled Refinement, and Primitive-Guided Implicit Intersection. Gaussian-Field Hybrid Surfaces use a hybrid approach combining Gauss surfaces and fields to reconstruct different interfaces. Transport-Coupled Refinement reduces coupling errors by evaluating the entire light path. Primitive-Guided Implicit Intersection combines primitive support regions to reduce errors in the intersection of interfaces and rays. In experiments on datasets such as RefRef, GaussTrans achieves significantly higher accuracy with relatively low spatial storage requirements.
Figures & tables
Figure 2: GaussTrans overview. (1–3) Inputs and local fitting construct Gaussian-Field Hybrid Surfaces. (4) Transport-Coupled Refinement updates interface geometry using multi-view path errors. (5) Primitive-Guided Implicit Intersection proposes and validates intersections. (6) Light transport through the interfaces produces rendered views.
Figure 3: Gaussian-Field Hybrid Surfaces. (a) Irregular local patches and their normals. (b) Position-dependent weighted least-squares interpretation of the field value. (c) The fused zero-level surface and its gradient normal. Colors identify local patches, not separate physical interfaces.
Figure 4: Layer count and budget. (a, b) Spherical/ellipsoidal shells, fixed per-interface capacity. (c) Ellipsoids, fixed total storage. Landing errors use 65 fixed primary rays/five views; dense source has higher storage. (d) With the inner surface held fixed, solid lines evenly distribute a fixed total storage budget across interfaces; dashed lines keep the per-interface representation capacity fixed. Units: scene units/degrees.
Figure 5: Six-shell accuracy and path completion under five approximately byte-matched Gaussian/QEM [ 10 ] budgets. (a) Mean landing error on the same 44 rays completed by all 15 configurations. (b) Complete 24-event paths/65; GaussTrans and smooth QEM both reach 65 at every budget. Storage: uncompressed numerical geometry. Errors: scene units.
(a) Local surface degree
Degree N
2
3
4
5
6
7
8
Shape coeffs.
3
7
12
18
25
33
42
Landing ( 10−3 )
2.564
1.498
0.851
0.616
0.637
0.488
0.507
Angle ( 10−2 deg)
4.189
2.428
1.394
1.011
1.045
0.794
0.825
Time (s)
8.149
8.046
9.706
11.144
13.036
14.913
17.078
Bank (KiB)
712.5
862.5
1050.0
1275.0
1537.5
1837.5
2175.0
Table 1: Local-degree and update ablations. Errors use scene units/degrees at the indicated scales; bold marks best means.
Method
PSNR ↑
SSIM ↑
LPIPS ↓
Object MAE ↓
Representation / asset scope
GaussTrans
23.49
0.9085
0.0665
0.1409
82.85 KiB (geometry)
Mesh (same inputs)
23.45
0.9068
0.0838
0.1414
2140.17 KiB (geometry)
RT-Splatting [ 35 ]
23.44
0.9010
0.0877
0.1442
125.77 MB (inference)
TransparentGS [ 15 ] a
22.16
0.8923
0.0953
0.1480
1.586M Gaussians
TNSR [ 6 ]
22.42
0.8741
0.1600
0.1235
67.78 MB (inference)
R3F [ 42 ] (estimated mesh)
24.54
0.8999
0.0925
0.1263
radiance field + mesh
Table 2: RefRef sphere: 100 test views at 512×512 ; metrics are averaged over views. Only ours and Mesh use the same inputs. KiB = 210 bytes, MB = 106 bytes; primitive counts follow different accounting conventions (Supplementary Sec. S1).
Method
Core geometry (KiB) ↓
Normal cache (KiB)
Common-pixel MAE bounds ↓
Hard pixels all ↓
Hard pixels four-crossing ↓
GaussTrans
96.88
0.00
[0.07490, 0.07534]
0
0
QEM matched (F)
96.82
0.00
[0.17436, 0.17541]
0
0
QEM matched (S)
96.82
32.27
[0.08844, 0.08896]
768
137
QEM double (S)
193.71
64.57
[0.06980, 0.07017]
382
69
Dense Mesh (F)
648.00
0.00
[0.11830, 0.11879]
0
0
Dense Mesh (S)
648.00
216.00
[0.04808, 0.04845]
166
28
Table 3: RefRef torus [ 42 ] ; QEM [ 10 ] : one 512×512 view, four-crossing region. Brackets: RGB MAE bounds from unresolved transport, not confidence intervals. All hard failures remain counted.