GaussianPile: A Unified Sparse Gaussian Splatting Framework for Slice-based Volumetric Reconstruction

Di Kong; Yikai Wang; Wenjie Guo; Yifan Bu; Boya Zhang; Yuexin Duan; Xiawei Yue; Wenbiao Du

cs.CVMar 21, 2026

GaussianPile: A Unified Sparse Gaussian Splatting Framework for Slice-based Volumetric Reconstruction

Authors: Di Kong, Yikai Wang, Wenjie Guo, Yifan Bu, Boya Zhang, Yuexin Duan, Xiawei Yue, Wenbiao Du, +3 more

Organizations: Tsinghua University · Zhongguancun Academy · Beijing Normal University · Nankai University · Beijing Institute of Technology · Beihang University

Abstract

Slice-based volumetric imaging is widely applied and it demands representations that compress aggressively while preserving internal structure for analysis. We introduce GaussianPile, unifying 3D Gaussian splatting with an imaging system-aware focus model to address this challenge. Our proposed method introduces three key innovations: (i) a slice-aware piling strategy that positions anisotropic 3D Gaussians to model through-slice contributions, (ii) a differentiable projection operator that encodes the finite-thickness point spread function of the imaging acquisition system, and (iii) a compact encoding and joint optimization pipeline that simultaneously reconstructs and compresses the Gaussian sets. Our CUDA-based design retains the compression and real-time rendering efficiency of Gaussian primitives while preserving high-frequency internal volumetric detail. Experiments on microscopy and ultrasound datasets demonstrate that our method reduces storage and reconstruction cost, sustains diagnostic fidelity, and enables fast 2D visualization, along with 3D voxelization. In practice, it delivers high-quality results in as few as 3 minutes, up to 11x faster than NeRF-based approaches, and achieves consistent 16x compression over voxel grids, offering a practical path to deployable compression and exploration of slice-based volumetric datasets.

Figures & tables

Figure 1: Panel (a) is the average PSNR-fps-minute comparison (circle radius encodes minutes of training). Our method achieves the highest accuracy at far lower compute cost than prior work. Panels (b–d) are the comparisons of Gaussian rendering models under different imaging physics. (b) All-in-focus ( $\sigma_{z}\to\infty$ ): no axial falloff; primitives contribute regardless of its depth, appropriate for all-in-focus pinhole rendering (e.g., original 3DGS [ 14 ] ) or line-integral modalities (e.g., X-ray [ 4 ] ). (c) Zero-thickness ( $\sigma_{z}\to 0$ ): delta-like axial response; primitives contribute only at the exact plane, suitable for dense slicing (e.g., MRI [ 20 ] ). (d) Finite-thickness Focus Gaussian (ours): finite axial sensitivity ( $\sigma_{z}$ finite); primitives contribute within a focal zone determined by $\sigma_{z}$ , modeling anisotropic PSF in slice-based modalities (e.g., ultrasound, light-sheet microscopy).

Figure 2: Focus-aware rendering pipeline of GaussianPile. Given 3D Gaussians and a sensitivity map defining the focal zone, the rendering process consists of three stages: (1) Scan : project Gaussians onto slices at different depths; (2) Axial re-parameterization and Opacity modulation : apply axial weighting to attenuate off-focal contributions, yielding Focus Gaussians, while modulating opacity based on distance from the focal plane; (3) Screen-space projection and Additive rasterization : compute 2D marginal distributions and additively accumulate weighted Gaussian footprints to form final rendered images.

Figure 3: Reconstruction results of our GaussianPile at different iterations, visualized on the 3D data of a noisy Tribolium castaneum embryo by confocal microscopy with high laser power, demonstrating high 3D fidelity throughout the optimization process.

Figure 4: Qualitative comparison of reconstruction results. Slice examples of different methods with PSNR (dB) shown at the bottom right of each image. The up three rows are from the TDSC-ABUS dataset [ 24 ] and the down three rows are from the rDL-LSM dataset [ 30 ] . Our method shows superior performance in modeling different regions (e.g. breast fibroglandular tissues and tumors) while preserving finer details compared to existing approaches.

Method	ABUS			rDL-LSM			TNNI1			hiPSC			Tribolium
Method	PSNR ↑	SSIM ↑	Time ↓	PSNR ↑	SSIM ↑	Time ↓	PSNR ↑	SSIM ↑	Time ↓	PSNR ↑	SSIM ↑	Time ↓	PSNR ↑	SSIM ↑	Time ↓
HEVC [ 34 ]	29.67	0.729	-	29.34	0.615	-	35.76	0.863	-	19.73	0.450	-	35.51	0.710	-
INIF [ 8 ]	24.84	0.608	1h24m	17.54	0.372	54m	30.89	0.826	27m	18.49	0.322	29m	31.55	0.517	27m
CoordNet [ 12 ]	23.32	0.517	60m	31.98	0.631	47m	30.12	0.827	32m	24.78	0.627	60m	29.42	0.687	46m
NeurComp [ 22 ]	19.85	0.341	2h41m	30.32	0.602	52m	39.25	0.915	11m	21.61	0.670	43m	32.59	0.719	21m
3DGS [ 14 ]	27.41	0.794	24m	28.63	0.749	10m	38.17	0.931	12m	20.15	0.522	26m	27.81	0.715	1h4m

Table 1: 2D quantitative comparison of GaussianPile and other methods.

Method	ABUS		rDL-LSM		TNNI1		hiPSC		Tribolium high		Tribolium low		Tribolium very low
Method	Mem. ↓	Comp. ↑	Mem. ↓	Comp. ↑	Mem. ↓	Comp. ↑	Mem. ↓	Comp. ↑	Mem. ↓	Comp. ↑	Mem. ↓	Comp. ↑	Mem. ↓	Comp. ↑
INIF [ 8 ]	7.8	15 $\times$	4.9	15 $\times$	1.8	18 $\times$	4.7	15 $\times$	3.1	16 $\times$	3.1	16 $\times$	3.1	16 $\times$
CoordNet [ 12 ]	10.7	11 $\times$	5.2	14 $\times$	3.2	10 $\times$	5.5	13 $\times$	5.6	9 $\times$	5.6	9 $\times$	5.6	9 $\times$
NeurComp [ 22 ]	7.4	16 $\times$	4.5	16 $\times$	2.0	16 $\times$	4.5	16 $\times$	3.0	16 $\times$	3.0	16 $\times$	3.0	16 $\times$
3DGS [ 14 ]	691.7	0.2 $\times$	50.2	1.5 $\times$	10.6	3 $\times$	173.6	0.4 $\times$	28.2	1.7 $\times$	378.4	0.1 $\times$	45.0	1.1 $\times$
Ours w/o Quant.	49.8	2 $\times$	22.6	3 $\times$	9.7	3 $\times$	21.8	3 $\times$	9.3	5 $\times$	11.3	4 $\times$	11.0	4 $\times$

Table 2: Compression comparison of GaussianPile and other methods. Memory size is reported in megabytes (MB).

Method	ABUS		rDL-LSM		TNNI1		hiPSC		Tribolum
Method	PSNR	SSIM	PSNR	SSIM	PSNR	SSIM	PSNR	SSIM	PSNR	SSIM
CoordNet [ 12 ]	18.75	0.454	32.03	0.689	29.26	0.716	22.23	0.642	28.56	0.693
NeurComp [ 22 ]	19.38	0.248	31.08	0.586	39.53	0.940	22.54	0.653	32.51	0.667
3DGS [ 14 ]	28.49	0.828	25.39	0.665	26.75	0.780	15.20	0.349	21.46	0.583
Ours	33.22	0.876	34.47	0.814	40.17	0.952	33.73	0.752	35.05	0.860

Table 3: 3D quantitative results of baselines and our method.

Figure 5: Results of 3DGS and our method with PSNR (dB) indicated on each image. We show slices of 3DGS queried from elevational angle (z-axis). Although 3DGS can produce plausible 2D projections, its reconstructed volume exhibits poor quality.

Tribolium high	PSNR ↑	SSIM ↑	Time ↓	Gau.
$\sigma_{z}=\delta_{z}/10$	34.25	0.927	8m9s	215k
$\sigma_{z}=\delta_{z}/2$	36.44	0.942	8m18s	214k
$\bm{\sigma_{z}=\delta_{z}}$	36.67	0.944	7m23s	211k
$\sigma_{z}=2\delta_{z}$	36.12	0.940	8m8s	162k
$\sigma_{z}=10\delta_{z}$	23.05	0.858	9m17s	66k
$M=50\text{k}$	34.12	0.932	7m28s	85k

Table 4: Ablation results with our choices in bold.

Figure 6: Qualitative comparison of reconstruction results on three cellular datasets. Slice examples of different methods with PSNR (dB) shown at the bottom of each image.

Figure 7: 3D reconstruction results on cellular datasets, presented by depth-encoded color rendering of volumes.

Figure 8: Convergence analysis of baseline methods and GaussianPile. We visualize the intermediate snapshots at different iterations. GaussianPile shows faster and better convergence.

Figure 9: Qualitative results on industrial CeraMIRScan MIR-OCT data, validating GaussianPile’s applicability to diverse slice-based imaging modalities.

Figure 10: Generalization to highly anisotropic ssEM (CREMI 1 , ISBI12 [ 1 ] , $10\times/12.5\times$ anisotropy respectively). We visualize three consecutive slices for each dataset.

Figure 11: Scalability analysis on the large-scale EPFL CA1 EM dataset [ 23 ] .

Method	ISBI12						CREMI
Method	2D PSNR $\uparrow$	2D SSIM $\uparrow$	3D PSNR $\uparrow$	3D SSIM $\uparrow$	CR $\uparrow$	Time $\downarrow$	2D PSNR $\uparrow$	2D SSIM $\uparrow$	3D PSNR $\uparrow$	3D SSIM $\uparrow$	CR $\uparrow$	Time $\downarrow$
CoordNet [ 12 ]	25.19	0.708	24.98	0.715	3 $\times$	51m	22.66	0.620	20.80	0.589	6 $\times$	2h34m
NeurComp [ 22 ]	24.45	0.695	25.19	0.728	8 $\times$	5m	19.81	0.428	17.62	0.333	10 $\times$	11h
Ours	28.79	0.815	28.98	0.881	8 $\times$	6m	29.50	0.831	28.27	0.752	10 $\times$	25m

Table 5: Quantitative comparison on highly anisotropic ssEM datasets. We report 2D slice fidelity, 3D volumetric fidelity, and compression/runtime metrics on ISBI12 [ 1 ] and CREMI 1 .

Explore similar work

Efficient Compression of Structured and Unstructured Volumes via Learned 3D Gaussian Representation
Jul 1, 2026Landon Dyken, Sharmistha Chakrabarti, Nathan Debardeleben +43DGS Compression 3D Gaussian Splatting
NanoGS: Training-Free Gaussian Splat Simplification
Mar 17, 2026Butian Xiong, Rong Liu, Tiantian Zhou +33DGS Compression 3D Gaussian Splatting
USC Institute for Creative Technologies · Texas A&M University
Gaussian Volumetric Representation for Efficient Shear-Warp Visualization
Jul 28, 2026Mayuri Mathur, Ojaswa SharmaScientific Visualization Medical Image Representation Learning
Indraprastha Institute of Information Technology Delhi · Indraprastha Institute of Information Technology Delhi, New Delhi, India

Jul 1, 2026J/K move · Enter open · S save

Efficient Compression of Structured and Unstructured Volumes via Learned 3D Gaussian Representation

Landon Dyken, Sharmistha Chakrabarti, Nathan Debardeleben, Steve Petruzza+3

3DGS Compression 3D Gaussian Splatting

Open paper AskPDF arXiv

Recent work has shown that implicit neural representations (INRs) can be trained to effectively compress structured and unstructured volume data, allowing for direct data querying with a reduced memory footprint. However, as existing INRs for unstructured volumes do not encode geometry, they require partial mesh storage for later sampling, limiting achievable compression. At the same time, novel view synthesis methods have shown that explicit collections of 3D Gaussians can be used to accurately visualize volume data. In this work, we introduce an explicit model for volume data compression based on 3D Gaussian primitives. We reinterpret collections of 3D Gaussians as an explicit representation of a scalar field and use a sampling strategy that reconstructs scalar values at spatial locations through weighted aggregation of intersecting Gaussians. We develop optimized CUDA-accelerated pipelines for structured and unstructured model sampling, loss functions that encourage accurate domain encoding by our models, and a novel sampling-error based densification strategy. Our explicit formulation naturally encodes domain geometry, eliminating the need for mesh storage in unstructured volumes and introducing significantly higher compression opportunities. Compared to existing INRs, we demonstrate that our explicit model achieves competitive reconstruction quality with significant training speedups on structured volumes, while markedly outperforming in all metrics on unstructured volumes.

Abstract

Explore similar work

Efficient Compression of Structured and Unstructured Volumes via Learned 3D Gaussian Representation

NanoGS: Training-Free Gaussian Splat Simplification

Gaussian Volumetric Representation for Efficient Shear-Warp Visualization