Counting Contributions, Not Views: Occlusion-Aware Densification for 3D Gaussian Splatting

Marcel Rogge; René Schuster; Didier Stricker

cs.CVOct 19, 2026

Counting Contributions, Not Views: Occlusion-Aware Densification for 3D Gaussian Splatting

Authors: Marcel Rogge, René Schuster, Didier Stricker

Organizations: University Kaiserslautern-Landau – RPTU · German Research Center for Artificial Intelligence – DFKI

Abstract

3D Gaussian Splatting grows a dense scene from a sparse initialization through adaptive density control (ADC), driven by per-Gaussian statistics gathered during rendering. These statistics pass a frustum-based visibility filter blind to occlusion: fully occluded Gaussians still count as observed, diluting the averaged gradient signal and suppressing densification precisely where geometry is occluded. We correct this with an occlusion-aware filter that restricts densification statistics to views in which a Gaussian actually contributes, and replace the criterion's raw view-count normalizer with an average-contribution one. Integrated into a StopThePop-based framework, the two together improve every image-quality metric over the occlusion-blind baseline on occlusion-heavy scenes (up to 6% lower LPIPS), at less than half the Gaussians a view-count normalizer requires. Under the longer schedules these large scenes require, the count stays bounded where most criteria diverge, and on smaller benchmarks our method stays competitive at a fraction of the competitors' count.

Figures & tables

Panel of Figure 2 — Figure 2: Occlusion-aware versus frustum-based visibility filter. Showing all Gaussians in front of a camera from the Mip-NeRF360 garden scene, which (a) pass the frustum test or (b) actually contribute to the rendered image. (c) shows their difference in red.

Method	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	GMSD $\downarrow$	Count $\downarrow$	Time $\downarrow$
Occlusion-blind count	25.52	0.818	0.352	0.117	0.557	22.3
Occlusion-aware count	25.75	0.834	0.319	0.103	2.567	36.6
PixelGS	25.11	0.829	0.312	0.105	6.607	67.0
PixelGS + DS	25.59	0.824	0.341	0.112	1.443	28.8
Ours	25.81	0.830	0.330	0.106	1.133	29.9

Table 1: Occlusion-aware filter and densification criteria on the occlusion-heavy Zip-NeRF scenes (Mip-NeRF360 counterpart in Tab. S12). The occlusion-blind baseline, then varying only the criterion: the occlusion-aware observation count, the PixelGS pixel-count-weighted mean without and with distance scaling (+ DS), and our average contribution. Counts are in millions of Gaussians; time in minutes.

	Method	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	GMSD $\downarrow$	Count $\downarrow$	Time $\downarrow$
$100$ k	Occlusion-aware count	25.90	0.842	0.292	0.095	4.792	178.5
	PixelGS	diverges ( $>16.8$ M Gaussians, training crashes)
	PixelGS + DS	26.35	0.843	0.302	0.094	2.843	129.5
	Ours	26.48	0.847	0.295	0.091	1.890	105.6
$210$ k	Occlusion-aware count	diverges on 2 of 4 scenes (alameda, nyc)
	PixelGS + DS	26.46	0.848	0.288	0.090	4.020	328.4

Table 2: Densification criteria on the Zip-NeRF scenes under longer schedules (

100

k and

210

k iterations; single runs). Runs that crash from over-densification are marked as divergent. Counts are in millions of Gaussians; time in minutes.

Panel of Figure 3 — Figure 3: Qualitative comparison on Zip-NeRF scenes (berlin, london, nyc; $30$ k schedule): StopThePop [ 20 ] , our re-implementations of PixelGS [ 31 ] and Grubert et al . [ 9 ] on the StopThePop base, and our full model. The crops show the areas marked with red boxes in the test view. The state-of-the-art methods remain blurry in these occlusion-heavy scenes, while ours reconstructs fine geometry.

	Dataset	Deep Blending				Mip-NeRF360				Tanks & Temples				Zip-NeRF
	Metric	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	G(M) $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	G(M) $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	G(M) $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	G(M) $\downarrow$
3DGS	PixelGS †	28.95	0.892	0.250	$\sim$ 4.6	27.53	0.822	0.191	$\sim$ 5.5	23.83	0.854	0.151	$\sim$ 4.5	N/A	N/A	N/A	N/A
3DGS	Grubert et al . †	29.63	0.902	0.240	$\sim$ 2.3	27.71	0.824	0.193	$\sim$ 3.3	24.18	0.860	0.149	$\sim$ 2.4	N/A	N/A	N/A	N/A
StopThePop	STP (Reset)	29.91	0.904	0.234	2.819	27.30	0.814	0.214	3.289	23.15	0.843	0.173	1.800	25.11	0.808	0.366	0.951
	PixelGS	29.54	0.901	0.234	4.392	27.37	0.822	0.190	5.328	23.27	0.852	0.147	4.175	25.15	0.811	0.360	1.980
	Grubert et al .	30.05	0.905	0.232	2.209	27.59	0.824	0.192	3.263	23.78	0.858	0.146	2.231	25.27	0.812	0.360	1.131

Table 3: Comparison on Deep Blending, Mip-NeRF360, Tanks & Temples and Zip-NeRF. We list results as reported by Grubert et al . [ 9 ] (

\dagger

) and RadSplat [ 19 ] (

\ddagger

). All other results are from our own experiments on the shared StopThePop (STP) base.

G

is Gaussians in millions; rows are grouped by base, with RadSplat and Zip-NeRF as upper-bound references (excluded from the color scale).

Components	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	GMSD $\downarrow$	Count $\downarrow$	Time $\downarrow$
Full model	25.81	0.830	0.330	0.106	1.133	29.9
+ clone	20.86	0.735	0.432	0.175	0.760	28.6
- GOF Grad	25.56	0.818	0.351	0.119	0.781	23.8
- long-axis split	25.80	0.833	0.321	0.103	1.635	29.6
- contrib prune	25.79	0.830	0.330	0.106	1.245	27.9
+ warm-up stats	25.81	0.830	0.330	0.106	1.134	26.7

Table 4: Ablation on the Zip-NeRF scenes: removing (

-

) or adding (

+

) one component relative to our full model. The catastrophic failure row (gray) is excluded from the color scale. Counts are in millions of Gaussians; time in minutes.

Appendix figures & tables28 assets

Supplementary material from the paper’s appendix.

Appendix

	Method	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	Count $\downarrow$
2DGS	2DGS	27.69	0.850	0.218	2.059
	+ filter	27.87	0.857	0.204	2.512
	+ filter + criterion ( $0.001$ )	over-densifies; out of memory
	+ filter + criterion ( $0.002$ )	28.04	0.865	0.182	3.682
	+ filter + criterion ( $0.003$ )	27.89	0.855	0.204	2.184
	+ filter + criterion ( $0.004$ )	27.67	0.845	0.225	0.910

Appendix

Table S1 : Both ports on the Mip-NeRF360 subset of the threshold sweep (single runs, averaged over the four scenes): the vanilla base, the occlusion-aware filter alone, and the full port across the densification-threshold sweep. Shading compares within each base. Counts are in millions of Gaussians.

		2DGS				3DGS
	Method	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	Count $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	Count $\downarrow$
alameda	Vanilla	20.90	0.689	0.456	0.630	22.28	0.728	0.400	1.216
	+ filter	18.95	0.661	0.480	3.222	22.40	0.740	0.380	2.654
	+ filter + crit.	20.13	0.684	0.457	1.443	22.46	0.736	0.388	1.806
berlin	Vanilla	25.79	0.865	0.359	0.556	27.10	0.886	0.326	1.024
	+ filter	25.48	0.870	0.343	3.039	27.39	0.895	0.308	3.142

Appendix

Table S2 : Both ports on the Zip-NeRF scenes (

30

k, single runs), at each port’s own densification threshold (

0.002

for 2DGS,

0.003

for 3DGS). Shading compares within each scene and base. Counts are in millions of Gaussians.

	Threshold	Densification	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	Count $\downarrow$
2DGS	$0.002$	with clone	28.14	0.867	0.181	2.947
		split-only	28.19	0.865	0.185	1.696
	$0.003$	with clone	28.02	0.859	0.202	1.441
		split-only	27.97	0.855	0.206	0.980
3DGS	$0.002$	with clone	28.85	0.878	0.157	6.455
		split-only	28.60	0.870	0.164	5.267

Appendix

Table S3 : Isolating the clone operation on both ports, on the Mip-NeRF360 subset (single runs, averaged over the four scenes). Shading compares within each base. Counts are in millions of Gaussians.

	Method	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	GMSD $\downarrow$	Count $\downarrow$	Time $\downarrow$
Zip	StopThePop	24.95	0.802	0.376	0.133	0.486	24.3
Zip	Baseline	25.52	0.818	0.352	0.117	0.557	22.3
Mip	StopThePop	27.01	0.809	0.227	0.102	1.711	25.6
Mip	Baseline	27.29	0.821	0.196	0.087	2.117	31.4

Appendix

Table S4 : StopThePop [ 20 ] versus our integrated baseline (StopThePop with the adopted densification components, still occlusion-blind) on the Zip-NeRF (top) and Mip-NeRF360 (bottom) scenes (3-run averages). Counts are in millions of Gaussians; time in minutes.

Components	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	GMSD $\downarrow$	Count $\downarrow$	Time $\downarrow$
StopThePop	24.95	0.802	0.376	0.133	0.486	24.3
+ contributions	23.03	0.776	0.393	0.148	1.099	26.6
+ extent	25.06	0.809	0.365	0.127	0.564	20.1
+ contributions + extent	25.55	0.824	0.341	0.114	1.303	27.0
+ split-only	25.19	0.808	0.366	0.129	0.570	19.7
+ contributions + split-only	25.63	0.822	0.343	0.115	1.212	25.2

Appendix

Table S5 : Complete clone-fix ablation on the Zip-NeRF scenes (3-run averages): StopThePop plus our two contributions and/or one clone fix—split-only densification or the manual scene-extent override of Sec. 0.C.1 . The catastrophic failure row (gray) is excluded from the color scale. Counts are in millions of Gaussians; time in minutes.

Panel of Figure 12 — Figure S1 : The clone failure mode on a nyc test view ( $30$ k schedule; one of the three ablation runs). Top row: StopThePop ( $25.5$ dB) reconstructs the view, then + contributions collapses it into scene-covering floater clusters ( $12.3$ dB). Bottom row: + split-only behaves like StopThePop ( $25.7$ dB), then + contributions + split-only resolves the view best ( $28.5$ dB). Insets (bottom): our contributions recover the individual keyboard keys and near-legible on-screen text that split-only densification alone leaves blurred.

Dataset	Occlusion	Scene range	Train views
Zip-NeRF [ 2 ]	$0.630$	$0.568$ – $0.700$	$1336$
Tanks & Temples [ 16 ]	$0.482$	$0.388$ – $0.577$	$241$
Mip-NeRF360 [ 1 ]	$0.194$	$0.090$ – $0.342$	$189$
Deep Blending [ 11 ]	$0.145$	$0.083$ – $0.208$	$214$

Appendix

Table S6 : Occlusion rate ( Eq. S6 ) per dataset, averaged over five trained models and all training views. The range gives the minimum and maximum over the dataset’s scenes. Training views are averages.

Dataset	Scene	Occlusion	LPIPS base	$\Delta$ LPIPS
Deep Blending	playroom	$0.083$	$0.2291$	$-0.0037$ ( $-1.6\%$ )
Mip-NeRF360	room	$0.089$	$0.2062$	$+0.0047$ ( $+2.3\%$ )
Mip-NeRF360	kitchen	$0.144$	$0.1253$	$+0.0074$ ( $+5.9\%$ )
Mip-NeRF360	bicycle	$0.146$	$0.1797$	$+0.0015$ ( $+0.8\%$ )
Mip-NeRF360	stump	$0.178$	$0.1910$	$+0.0062$ ( $+3.3\%$ )
Mip-NeRF360	garden	$0.179$	$0.1029$	$+0.0063$ ( $+6.1\%$ )

Appendix

Table S7 : Per-scene occlusion rate against the LPIPS improvement of our two contributions over the occlusion-blind baseline.

\Delta

LPIPS is the improvement of ours. Sorted by occlusion rate.

	Method	G(M)	Occluded	Share
Zip	StopThePop (Reset)	0.951	149k	15.7%
	StopThePop (Decay)	0.486	50.9k	10.5%
	PixelGS [ 31 ]	1.980	449k	22.7%
	Grubert et al . [ 9 ]	1.131	156k	13.8%
	Ours (no contribution prune)	1.245	89.1k	7.2%
	Ours (Reset)	1.991	0	0%

Appendix

Table S8 : Fully occluded Gaussians (those that never contribute to any training view) in each method’s reported model, on the Zip-NeRF (top) and Mip-NeRF360 (bottom) scenes (3-run averages). The occlusion-blind methods carry substantial dead weight, up to

23\%

on the occlusion-heavy scenes. Our contribution prune removes them during training, so the final models contain none.

Threshold	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	GMSD $\downarrow$	Count
$0.001$	27.98	0.864	0.162	0.077	4.90
$0.0015$	28.25	0.874	0.157	0.073	2.75
$\mathbf{0.002}$	28.31	0.875	0.160	0.073	1.79
$0.0025$	28.24	0.874	0.166	0.074	1.30
$0.004$	28.16	0.865	0.191	0.076	0.66

Appendix

Table S9 : Threshold calibration of our criterion on the Mip-NeRF360 calibration subset (bicycle, bonsai, counter, garden), single runs, using our full integrated configuration: the occlusion-aware filter, the average-contribution criterion, and the adopted densification components. Our operating point

0.002

(bold) attains the best PSNR, SSIM, and GMSD of the sweep; over- and under-densification degrade quality on either side. Counts are in millions of Gaussians.

Floor	Threshold	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	GMSD $\downarrow$	Count
$0$	$0.002$	diverges on $4/4$ scenes; peak $17.7$ – $21.0$ M by iteration ${\sim}1{,}800$
$0$	$0.004$	diverges on $4/4$ scenes; peak $17.4$ – $20.1$ M by iteration ${\sim}2{,}300$
$0$	$0.01$	diverges on $4/4$ scenes; peak $17.2$ – $17.9$ M by iteration ${\sim}5{,}000$
$0$	$0.02$	diverges on $3/4$ scenes; berlin completes at $2.00$ M
$0$	$0.03$	$25.15$	$0.803$	$0.369$	$0.123$	$1.63$
$0$	$0.05$	$24.36$	$0.775$	$0.417$	$0.150$	$0.21$

Appendix

Table S10 : Floor

0

across a threshold ladder on the Zip-NeRF scenes (

30

k, single runs). Diverging rows report how many scenes crossed the trainable limit, together with the peak count reached and the iteration by which training stopped. The last two rows give floor

1

for reference: our operating point, and the setting whose count is closest to the only floor-

0

threshold that completes. Counts are in millions of Gaussians.

Floor	Threshold	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	GMSD $\downarrow$	Count
$0.5$	$0.002$	$25.83$	$0.836$	$0.314$	$0.100$	$2.39$
$1$	$0.002$	$25.80$	$0.830$	$0.330$	$0.106$	$1.13$
$2$	$0.002$	$25.58$	$0.818$	$0.353$	$0.117$	$0.52$
$4$	$0.002$	$25.11$	$0.801$	$0.381$	$0.132$	$0.24$
$1$	$0.004$	$25.44$	$0.812$	$0.363$	$0.121$	$0.40$
$0.5$	$0.004$	$25.71$	$0.824$	$0.342$	$0.111$	$0.78$

Appendix

Table S11 : The denominator floor on the Zip-NeRF scenes (

30

k, single runs). Upper block: at the default threshold, the floor acts as a count–quality trade off. Lower block: with each floor’s threshold scaled by a predicted value, interleaved with a threshold sweep of floor

1

and ordered by count. Counts are in millions of Gaussians.

Method	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	GMSD $\downarrow$	Count $\downarrow$	Time $\downarrow$
Occlusion-blind count	27.29	0.821	0.196	0.087	2.117	31.4
Occlusion-aware count	27.33	0.822	0.188	0.086	2.595	35.4
PixelGS	27.22	0.819	0.184	0.087	3.433	43.8
PixelGS + DS	27.20	0.820	0.183	0.087	3.157	42.2
Ours	27.30	0.820	0.186	0.087	1.954	35.4

Appendix

Table S12 : Densification criteria on Mip-NeRF360 (same criteria as Tab. 1 of the main paper; 3-run averages) and the occlusion-blind count of the integrated framework for reference. Counts are in millions of Gaussians; time in minutes.

	Method	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	GMSD $\downarrow$	Count $\downarrow$
Full	StopThePop	27.01	0.809	0.227	0.102	1.711
	Occlusion-blind count	27.29	0.821	0.196	0.087	2.117
	Occlusion-aware count	27.33	0.822	0.188	0.086	2.595
	Ours	27.30	0.820	0.186	0.087	1.954
Indoor	StopThePop	30.16	0.917	0.194	0.064	0.756
	Occlusion-blind count	30.60	0.922	0.182	0.060	0.741

Appendix

Table S13 : Mip-NeRF360 performance split by scene type (3-run averages): the full 9-scene average, the 4 indoor scenes (bonsai, counter, kitchen, room), and the 5 outdoor scenes (bicycle, flowers, garden, stump, treehill). Counts are in millions of Gaussians.

	Flower bed				Grass ring
Criterion	Count	PSNR $\uparrow$	SSIM $\uparrow$	GMSD $\downarrow$	Count	PSNR $\uparrow$	SSIM $\uparrow$	GMSD $\downarrow$
Blind	0.26	20.52	0.694	0.169	0.18	19.49	0.279	0.216
OA-count	0.42	20.77	0.712	0.163	1.80	19.42	0.265	0.202
Ours	0.97	20.71	0.719	0.159	0.62	18.95	0.243	0.205

Appendix

Table S14 : Region-based allocation and image quality on the flowers scene (3-run averages). The two regions are the flower bed and the surrounding grass ring (radius

<1.9

and

1.9

–

2.9

from the centre of the flower bed in the top-down projection, respectively). For each region, Count is the number of Gaussians (in millions) positioned within it, and PSNR/SSIM/GMSD are measured on rough image crops of those regions across all test views.

Figure S2 : Robustness to the densification threshold on the Mip-NeRF360 subset: image quality versus final Gaussian count. Each data point represents a different threshold decreasing from left to right, as the count increases. All points are three-run averages.

	Method	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	GMSD $\downarrow$	Count $\downarrow$
Full	StopThePop	24.67	0.813	0.372	0.152	0.396
Full	Ours	25.58	0.833	0.344	0.128	0.832
Half	StopThePop	24.95	0.802	0.376	0.133	0.486
Half	Ours	25.81	0.830	0.330	0.106	1.133

Appendix

Table S15 : Zip-NeRF trained with full-resolution images against the half-resolution protocol used throughout the paper (

30

k, single runs). Shading compares within each resolution block. Counts are in millions of Gaussians.

Method	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	GMSD $\downarrow$	Count $\downarrow$	Time $\downarrow$
Occlusion-blind count	25.52	0.818	0.352	0.117	0.557	22.3
Occlusion-aware count	25.75	0.834	0.319	0.103	2.567	36.6
PixelGS	25.11	0.829	0.312	0.105	6.607	67.0
PixelGS + DS	25.59	0.824	0.341	0.112	1.443	28.8
PixelGS-ours hybrid	25.18	0.828	0.313	0.104	6.366	65.2
Ours	25.81	0.830	0.330	0.106	1.133	29.9

Appendix

Table S16 : The PixelGS-ours hybrid on the occlusion-heavy Zip-NeRF scenes, added to the criterion comparison of the main paper (calibrated threshold

0.0003

, three runs), preceded by the occlusion-blind count of the integrated baseline; the extended-schedule row reports single runs. Counts are in millions of Gaussians; time in minutes.

Normalizer	Threshold	PSNR $\uparrow$	LPIPS $\downarrow$	GMSD $\downarrow$	Count
per-view	$0.002$	$23.80$	$0.431$	$0.162$	$0.06$
	$0.0005$	$25.33$	$0.366$	$0.125$	$0.45$
	$0.00025$	$25.74$	$0.329$	$0.107$	$1.60$
sum-allocation	$0.0003$	$22.86$	$0.455$	$0.176$	$0.12$
	$0.0002$	$22.93$	$0.449$	$0.173$	$0.89$
	$0.0001$	diverges on $3/4$ scenes; berlin completes at $1.79$ M

Appendix

Table S17 : Alternative normalizers on the Zip-NeRF scenes (

30

k, single runs), swept to bring each into our count range; our operating point is the

3

-run average of the main paper. Counts are in millions of Gaussians.

Figure S3 : Training-camera distribution on the alameda scene (top-down). Dots mark the COLMAP camera positions, colored by the local density of training cameras; the sparse reconstruction is shown underneath for orientation. Interior hotspots are observed by an order of magnitude more cameras than the garden loop or the peripheral rooms.

Method	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	GMSD $\downarrow$	Count $\downarrow$	Time $\downarrow$
PixelGS + DS	26.46	0.848	0.285	0.090	4.472	553.5
Ours	26.59	0.852	0.278	0.086	2.754	395.7

Appendix

Table S18 : Extended

300

k schedule on the Zip-NeRF scenes (single runs), continuing the main paper’s extended-training comparison. Counts are in millions of Gaussians; time in minutes.

Figure S4 : Gaussian count versus iteration on the four Zip-NeRF scenes (single run per scene,

210

k solid,

300

k dashed, dotted verticals mark each schedule’s densification end, crosses mark crashes at the trainable limit). Ours (green) and PixelGS+DS (orange) stay bounded. The PixelGS criterion (purple) itself diverges on every scene, while the occlusion-aware count (red) diverges on two scenes.

Panel of Figure 29 — Figure S5 : Distance scaling (DS) suppresses near-object densification (alameda, $300$ k schedule). On this close-up test view, PixelGS with DS ( $6.1$ M Gaussians) leaves the near object a featureless blur, whereas our average-contribution criterion ( $3.1$ M Gaussians) reconstructs the detail well.

Panel of Figure 30 — Figure S6 : An under-observed corner view (alameda test view 00070, $300$ k schedule). The region is not intrinsically difficult but instead under-observed. The distance-scaled PixelGS renders an inoffensive smooth blur. Ours densifies to reconstruct the pixel-level color artifacts present in the ground truth, which it renders as colorful speckle.

Figure S7 : Per-view PSNR distribution per criterion across four training schedules. Numbers below each box count the views under

18

dB. The

30

k results are three-run averages; the extended schedules are single runs. The

210

k panel restricts all criteria to berlin and london, the two scenes on which all criteria complete.

Panel of Figure 32 — Figure S8 : The three teaser locations of the main paper (Fig. 1) as full test frames (ground truth, StopThePop, ours). Frame colors match the location boxes of the teaser.

Panel of Figure 33 — Figure S9 : Full-frame renders and error maps against vanilla PixelGS [ 31 ] on three Zip-NeRF scenes ( $30$ k, single runs). Columns: ground truth, PixelGS, ours, and the gradient-magnitude-similarity error map ( $1-$ GMS, brighter is worse) of each. For every scene, the first row is the test view on which our LPIPS advantage over PixelGS is largest and the second the view on which it is smallest. Row labels give the per-view GMSD as PixelGS $\rightarrow$ ours.

Figure S10 : Top-down Gaussian density of two Mip-NeRF360 scenes (log Gaussians per ground cell; the color scale is shared across the three criteria within each scene), comparing the occlusion-blind count (Blind), the occlusion-aware count (OA-count), and our average-contribution criterion (Ours) on the same integrated baseline. Camera positions are blue dots, and each map is annotated with its total Gaussian count. Kitchen (top): Ours concentrates Gaussians on the well-observed central object. Flowers (middle): OA-count over-densifies the flat lawn around the flower bed (its bright disk balloons outward), whereas Ours keeps the bed dense and the lawn lean. Bottom: render crops of a flowers test view, the flower bed (top) and the surrounding grass (bottom), each as Blind $|$ OA-count $|$ Ours, followed by two difference maps relative to Blind (OA-count $-$ Blind and Ours $-$ Blind), red denser. The former isolates the occlusion-aware filter, which adds Gaussians across the bed and the lawn; the latter shows the full effect of Ours, which concentrates them on the observed bed. Blind renders visibly blurrier grass, while OA-count and Ours are similarly sharp ( Tab. S14 ).

3D Gaussian Splatting

3D Gaussian Splatting represents scenes as optimizable 3D Gaussian primitives for fast, photorealistic rendering from new viewpoints.

See all 1,019 papers

OP2GS: Object-Aware 3D Gaussian Splatting with Dual-Opacity Primitives
May 19, 2026Guiyu Liu, Niklas Vaara, Janne Mustaniemi +23D Gaussian Splatting Open-Vocabulary 3D Segmentation
Center for Machine Vision and Signal Analysis, University of Oulu, Finland · Aalto University, Finland
DensiTok: Making Feed-Forward 3D Gaussian Splatting See More Views Than It Is Given
Oct 6, 2026Minhyeok Lee, Jungho Lee, Minseok Kang +33D Gaussian Splatting Sparse-View 3D Reconstruction
Yonsei University · NAVER AI Lab · Korea Institute of Science and Technology (KIST)
AVSplat: Dense-View Feed-Forward 3D Gaussian Splatting with Assist-View Preconditioning
Sep 5, 2026Muyu Xu, Fangneng Zhan, Yu Wei +23D Gaussian Splatting Novel View Synthesis
Nanyang Technological University, Singapore 639798, Singapore · Harvard University, MA 02138, USA

May 19, 2026J/K move · Enter open · S save

OP2GS: Object-Aware 3D Gaussian Splatting with Dual-Opacity Primitives

Guiyu Liu, Niklas Vaara, Janne Mustaniemi, Juho Kannala+1

Center for Machine Vision and Signal Analysis, University of Oulu, Finland · Aalto University, Finland

3D Gaussian Splatting Open-Vocabulary 3D Segmentation 3D Scene Representation 3D Instance Segmentation 3DGS Segmentation

Open paper AskPDF arXiv

3D Gaussian Splatting (3DGS) provides an explicit and efficient scene representation, but its primitives lack inherent object-level identity, hindering downstream tasks such as open-vocabulary scene understanding. Existing methods typically address this by either distilling high-dimensional feature embeddings into Gaussians or by lifting 2D mask labels into 3D via heuristic refinement. However, feature-based approaches incur heavy storage and decoding overhead, while lifting-based pipelines remain vulnerable to label contamination: Gaussians necessary for appearance reconstruction often receive incorrect object labels during 2D-to-3D projection. We propose OP2GS, an object-aware Gaussian representation that augments each primitive with an explicit instance identity and a dedicated instance opacity

σ^{*}

for object-mask rendering. The original opacity

σ

remains responsible for visual reconstruction, while

σ^{*}

models whether a Gaussian should contribute to a particular object mask. This dual-opacity formulation decouples visual existence from instance occupancy: mislabeled Gaussians can remain available for image rendering while becoming transparent in the object-mask branch. To learn this representation, we introduce a random object loss that optimizes the 1D instance occupancy field using the standard transmittance-based visibility of 3DGS. Semantic descriptors are then attached at the object level through multi-view aggregation, eliminating per-Gaussian feature storage. Compared with feature-training approaches, OP2GS achieves competitive open-vocabulary performance while significantly reducing computational overhead. Compared with training-free pipelines, it leverages physically consistent occupancy learning to resolve visibility ambiguities.

Abstract

3D Gaussian Splatting

Explore similar work

OP2GS: Object-Aware 3D Gaussian Splatting with Dual-Opacity Primitives

DensiTok: Making Feed-Forward 3D Gaussian Splatting See More Views Than It Is Given

AVSplat: Dense-View Feed-Forward 3D Gaussian Splatting with Assist-View Preconditioning