PKU-GS

A Large-Scale UAV-Based Real-World Dataset and Benchmark for 3D Gaussian Splatting Compression

Large-ScaleUAV CaptureCompression Benchmark

Dataset Overview

Dataset at a Glance

PKU-GS presents key quantitative indicators and visual summaries to highlight the dataset’s scale, coverage, and data collection characteristics.

Donut chart visualizing the dataset composition summary. — Distribution of PKU-GS scene categories.

188

Scenes

8 Million

Frames

1500+ km

Distance

> 8 km²

Coverage

Representative Scenes

A selection of representative scenes from PKU-GS

The following representative scenes demonstrate the wide range of spatial layouts, structural patterns, and visual complexity covered by PKU-GS.

A UAV view over a dense heritage courtyard block with detailed roof structures.

Historic Architecture

Compact roof geometry, repeated structural patterns, and narrow corridors create rich and challenging high-frequency details.

S-01

Historic Courtyard Block

Historic Architecture

Dense RooflinesOblique UAVFine Detail

A UAV view of a campus library tower surrounded by trees and institutional buildings.

Institutional Scene

The tall building facade and surrounding tree canopy create a mix of planar surfaces, depth variation, and partial occlusion.

S-02

Campus Library Tower

Institutional Scene

CampusVegetationFacade Geometry

A UAV view of a market edge with parking lots, storefronts, and rich visual clutter.

Urban Commercial Scene

Parked vehicles, signage, open paved areas, and irregular storefront boundaries increase scene heterogeneity.

S-03

Market Edge Zone

Urban Commercial Scene

Parking AreaStreet ClutterSignage

A UAV view of a large academic complex positioned above a waterside boundary.

Large-Scale Campus

Long-span structures and water-adjacent geometry make this scene valuable for evaluating large-scale consistency and compression performance.

S-04

Waterside Academic Complex

Large-Scale Campus

Water EdgeLong SpanMulti-Level

A UAV view of a contemporary campus with layered building volumes and road networks.

Contemporary Architecture

Contemporary glass-dominant architecture and layered building volumes provide a representative modern urban scene for capture and evaluation.

S-05

Research and Innovation Campus

Contemporary Architecture

Modern FacadesRoad NetworkLayered Volumes

A UAV view of a stadium bowl and sports field surrounded by tiered seating and urban context.

Sports Venue

Large-span roof geometry, an open infield structure, and repetitive seating elements make this scene a distinctive benchmark for sports venues.

S-06

Athletics Field

Sports Venue

GrandstandOpen FieldCurved Roof

Capture & Processing Pipeline

Pipeline from UAV Acquisition to Compression Evaluation

The figure summarizes the end-to-end workflow from UAV data acquisition to 3D Gaussian Splatting reconstruction, rendering, and evaluation.

Overview diagram of the PKU-GS capture and processing pipeline, from UAV data capture to video preprocessing, camera pose estimation, and 3DGS reconstruction, rendering, and evaluation. — End-to-end PKU-GS pipeline covering UAV capture, preprocessing, pose estimation, and 3D Gaussian Splatting reconstruction and evaluation.

Step 1

UAV Capture

Long-range UAV trajectories collect large-scene observations with practical aerial viewpoints.

Step 2

Video Preprocessing

Raw sequences are organized, filtered, and prepared for structured downstream reconstruction.

Step 3

Pose Estimation

Camera poses are recovered to establish geometric consistency across frames and viewpoints.

Step 4

Gaussian Splatting Reconstruction

Scene representations are reconstructed for rendering, analysis, and compression-oriented study.

Step 5

Compression Benchmark

Methods are compared with quality, efficiency, and scene-level robustness in mind.

Benchmark

Benchmark Summary

A quantitative benchmark summary organized by scene and metric for direct comparison across methods.

Table 3: Quantitative comparison in terms of PSNR, SSIM, LPIPS, and model size. The overall rank is obtained by averaging per-scene ranks over four metrics: PSNR, SSIM, LPIPS, and model size. Higher PSNR and SSIM, and lower LPIPS and model size, indicate better performance.

Method	Rank	1-23				1-42				1-71				1-73				1-78				1-100				2-6				2-8				3-7				3-11				4-10				4-25				4-28				5-6				5-13				6-7				6-22				6-26				6-31				6-44
Method	Rank	PSNR	SSIM	LPIPS	Size [MB]	PSNR	SSIM	LPIPS	Size [MB]	PSNR	SSIM	LPIPS	Size [MB]	PSNR	SSIM	LPIPS	Size [MB]	PSNR	SSIM	LPIPS	Size [MB]	PSNR	SSIM	LPIPS	Size [MB]	PSNR	SSIM	LPIPS	Size [MB]	PSNR	SSIM	LPIPS	Size [MB]	PSNR	SSIM	LPIPS	Size [MB]	PSNR	SSIM	LPIPS	Size [MB]	PSNR	SSIM	LPIPS	Size [MB]	PSNR	SSIM	LPIPS	Size [MB]	PSNR	SSIM	LPIPS	Size [MB]	PSNR	SSIM	LPIPS	Size [MB]	PSNR	SSIM	LPIPS	Size [MB]	PSNR	SSIM	LPIPS	Size [MB]	PSNR	SSIM	LPIPS	Size [MB]	PSNR	SSIM	LPIPS	Size [MB]	PSNR	SSIM	LPIPS	Size [MB]	PSNR	SSIM	LPIPS	Size [MB]
HAC++	2.500	26.421	0.860	0.173	19.976	27.824	0.831	0.231	7.218	28.961	0.895	0.144	15.226	28.340	0.880	0.150	17.702	27.140	0.865	0.175	17.227	28.998	0.912	0.118	16.787	28.508	0.893	0.119	21.032	28.489	0.849	0.292	8.522	28.772	0.860	0.187	17.473	27.757	0.852	0.196	14.001	28.756	0.865	0.186	17.419	27.471	0.930	0.099	19.513	28.311	0.890	0.293	6.946	27.981	0.915	0.136	20.859	30.128	0.923	0.124	15.245	31.520	0.947	0.087	6.695	32.732	0.955	0.059	17.582	28.424	0.886	0.160	19.829	27.870	0.872	0.152	21.325	27.705	0.900	0.123	16.324
PC-GS	6.333	24.245	0.757	0.298	13.737	25.791	0.748	0.353	4.891	26.997	0.836	0.247	11.044	26.345	0.812	0.259	12.175	23.092	0.739	0.322	12.033	26.335	0.838	0.237	11.549	26.259	0.823	0.221	15.133	27.289	0.843	0.315	5.981	26.971	0.788	0.296	12.161	25.859	0.767	0.303	9.326	26.363	0.760	0.320	11.840	25.070	0.854	0.206	14.998	27.324	0.886	0.300	5.040	25.119	0.833	0.255	16.771	27.945	0.861	0.238	10.661	30.195	0.928	0.169	4.724	30.132	0.904	0.150	12.965	26.156	0.799	0.283	13.790	24.630	0.743	0.302	14.730	25.763	0.839	0.226	11.387
CAT	6.667	23.686	0.728	0.275	14.402	25.694	0.735	0.315	5.037	26.913	0.828	0.194	11.308	25.642	0.786	0.220	14.333	23.404	0.746	0.280	14.350	26.291	0.830	0.175	13.114	25.721	0.791	0.183	15.436	27.268	0.838	0.232	5.879	26.765	0.770	0.271	12.530	25.677	0.754	0.264	11.353	26.293	0.738	0.269	13.082	23.376	0.782	0.203	13.623	27.468	0.886	0.184	5.261	24.542	0.801	0.181	17.670	27.575	0.840	0.162	10.881	30.646	0.929	0.081	5.547	29.760	0.890	0.098	14.817	25.747	0.774	0.235	14.568	24.662	0.735	0.251	16.607	25.699	0.830	0.166	13.550
contextgs	6.833	25.082	0.794	0.256	29.602	27.056	0.802	0.287	10.721	27.465	0.851	0.228	19.089	27.239	0.834	0.234	24.362	24.904	0.800	0.260	29.260	27.752	0.871	0.196	30.165	27.203	0.851	0.188	26.773	27.634	0.852	0.304	9.615	28.104	0.828	0.249	23.656	26.962	0.811	0.258	18.385	27.387	0.814	0.260	22.572	26.330	0.889	0.166	29.372	27.602	0.894	0.289	8.652	26.488	0.876	0.195	38.089	28.648	0.881	0.215	17.806	32.224	0.947	0.137	12.997	30.892	0.917	0.130	22.025	27.628	0.846	0.233	27.485	25.407	0.781	0.260	38.916	27.057	0.874	0.184	23.152
HAC	7.333	24.547	0.769	0.286	25.350	26.216	0.766	0.333	8.094	27.299	0.843	0.236	19.453	26.147	0.808	0.263	22.162	24.833	0.793	0.279	21.749	26.041	0.835	0.234	21.692	26.856	0.841	0.201	26.839	27.644	0.852	0.304	11.471	27.477	0.803	0.279	21.779	26.233	0.782	0.288	16.948	26.940	0.784	0.294	20.873	25.520	0.867	0.193	23.348	27.598	0.892	0.292	9.387	25.552	0.849	0.234	27.089	28.472	0.875	0.222	18.358	30.735	0.934	0.157	8.237	30.795	0.915	0.134	23.233	26.631	0.815	0.266	23.551	25.102	0.762	0.283	26.443	26.142	0.850	0.213	19.865
RDO-gaussian	8.833	23.617	0.737	0.310	14.056	24.851	0.717	0.391	5.413	25.606	0.779	0.324	6.653	25.019	0.774	0.298	8.611	23.908	0.762	0.310	11.242	26.099	0.824	0.253	13.383	25.137	0.785	0.264	14.721	26.034	0.813	0.348	2.728	24.857	0.720	0.367	6.012	25.062	0.733	0.332	8.510	24.683	0.669	0.406	6.356	24.471	0.841	0.204	27.043	25.424	0.864	0.320	2.975	24.734	0.821	0.254	31.878	26.465	0.804	0.298	7.562	28.492	0.896	0.209	4.634	28.205	0.859	0.203	12.535	25.113	0.764	0.306	14.842	24.250	0.730	0.311	13.389	24.528	0.806	0.255	12.260
Scaffold-gs	9.667	25.282	0.813	0.229	287.701	28.609	0.865	0.206	253.848	27.848	0.867	0.198	254.261	27.442	0.848	0.212	275.421	24.626	0.805	0.253	286.129	28.394	0.889	0.166	287.151	27.208	0.859	0.174	250.227	27.872	0.862	0.289	144.619	27.648	0.836	0.235	247.447	27.140	0.827	0.236	194.238	27.206	0.829	0.237	202.018	26.160	0.893	0.157	235.048	27.677	0.898	0.281	132.520	26.546	0.885	0.181	303.557	28.866	0.891	0.198	174.855	33.142	0.956	0.120	182.918	31.665	0.934	0.101	227.471	27.896	0.866	0.202	255.769	25.676	0.801	0.243	319.983	26.782	0.878	0.173	225.657
compact3d	10.333	23.930	0.752	0.305	28.850	25.731	0.759	0.354	16.710	26.123	0.806	0.296	16.790	25.595	0.795	0.285	20.170	24.268	0.773	0.310	21.480	26.528	0.841	0.242	25.880	25.676	0.808	0.248	33.560	26.797	0.836	0.326	9.290	25.673	0.754	0.341	16.420	25.848	0.771	0.306	22.320	25.533	0.734	0.354	19.950	24.828	0.856	0.206	50.320	25.872	0.879	0.307	8.510	25.239	0.841	0.248	58.440	27.597	0.858	0.247	25.990	30.827	0.933	0.157	17.930	29.208	0.891	0.170	29.810	25.827	0.792	0.299	31.100	24.723	0.748	0.308	26.040	25.088	0.828	0.240	24.820
EAGLES	10.333	24.641	0.784	0.265	93.141	26.919	0.808	0.287	52.729	27.096	0.840	0.242	57.771	26.461	0.827	0.242	69.945	25.304	0.820	0.251	94.148	27.836	0.876	0.190	120.135	26.771	0.844	0.203	116.829	27.246	0.851	0.308	34.618	26.671	0.789	0.298	46.332	26.856	0.813	0.258	81.295	26.174	0.771	0.309	57.938	25.805	0.881	0.169	145.151	26.356	0.890	0.293	36.197					28.206	0.874	0.223	67.859	31.580	0.944	0.137	57.220	30.717	0.918	0.130	98.422	27.312	0.845	0.236	122.342	25.925	0.808	0.245	107.934	26.261	0.869	0.187	108.084
reduced3dgs	10.500	25.349	0.810	0.238	325.088	27.362	0.816	0.277	151.595	27.420	0.850	0.226	188.577	26.955	0.838	0.228	194.613	25.847	0.832	0.237	267.509	28.355	0.887	0.175	350.452	27.657	0.866	0.178	358.461	27.631	0.857	0.297	92.745	27.299	0.804	0.277	134.585	27.381	0.829	0.237	211.414	26.851	0.795	0.287	160.170	27.461	0.920	0.124	576.099	26.820	0.896	0.284	109.507	27.952	0.913	0.146	688.216	29.295	0.900	0.192	247.010	32.409	0.949	0.129	155.277	32.345	0.940	0.099	364.637	28.007	0.860	0.217	332.755	26.562	0.825	0.225	299.771	26.936	0.885	0.166	307.025
3DGS	10.500	25.445	0.813	0.234	1095.680	28.403	0.851	0.229	531.830	27.613	0.855	0.219	634.020	27.247	0.844	0.220	759.940	26.064	0.838	0.229	1046.528	28.615	0.891	0.169	1280.000	27.759	0.868	0.174	1300.480	27.734	0.859	0.294	365.568	27.514	0.810	0.270	524.288	27.572	0.834	0.231	829.440	27.024	0.800	0.279	626.688	27.560	0.921	0.122	1986.560	26.983	0.898	0.281	409.600	28.008	0.914	0.144	2293.760	29.400	0.902	0.189	886.860	32.600	0.951	0.127	590.848	32.457	0.942	0.096	1259.520	28.248	0.865	0.211	1269.760	26.768	0.831	0.218	1136.640	27.081	0.889	0.162	1157.120
gsplat	10.833	25.355	0.810	0.219	1208.320	28.285	0.847	0.216	717.824	27.423	0.849	0.213	667.648	27.116	0.841	0.207	781.312	26.019	0.838	0.226	1085.440	28.665	0.891	0.144	1392.640	27.628	0.865	0.145	1310.720	27.643	0.857	0.306	349.184	27.301	0.804	0.284	509.952	27.462	0.830	0.211	813.056	26.816	0.794	0.267	608.256	27.520	0.921	0.100	2129.920	26.745	0.896	0.292	436.224	28.010	0.915	0.104	2539.520	29.272	0.899	0.155	944.128	32.454	0.950	0.083	601.088	32.402	0.941	0.070	1341.440	28.198	0.863	0.194	1300.480	26.760	0.832	0.198	1126.400	26.965	0.887	0.139	1157.120
TC-GS	10.833	23.718	0.721	0.342	17.948	26.813	0.782	0.323	17.257	26.317	0.805	0.297	13.906	26.330	0.801	0.278	19.160	23.412	0.737	0.338	17.707	26.431	0.826	0.261	18.429	24.845	0.745	0.322	15.670	27.284	0.838	0.323	8.773	26.117	0.748	0.343	16.243	25.058	0.721	0.357	13.574	25.988	0.712	0.383	18.599	23.229	0.777	0.296	18.699	27.304	0.885	0.302	7.327	23.180	0.737	0.370	19.758	27.101	0.817	0.298	14.579	29.056	0.897	0.236	5.601	29.067	0.865	0.208	18.771	25.655	0.768	0.319	21.651	24.147	0.709	0.339	19.725	25.231	0.810	0.270	16.613
compact3dgs	11.833	24.033	0.763	0.285	66.946	26.079	0.774	0.326	45.829	26.132	0.809	0.280	44.641	25.628	0.800	0.267	56.018	24.379	0.787	0.281	59.623	26.519	0.848	0.215	76.307	25.748	0.816	0.225	86.120	26.814	0.835	0.322	33.982	26.088	0.770	0.311	45.810	25.599	0.769	0.291	55.082	25.401	0.745	0.319	51.399	23.719	0.846	0.208	135.339	25.977	0.875	0.311	32.366	25.097	0.852	0.217	146.077	26.906	0.844	0.248	57.836	29.386	0.916	0.167	50.467	29.399	0.902	0.142	83.908	25.766	0.802	0.263	87.527	24.776	0.765	0.275	68.281	25.171	0.839	0.210	75.081
Minisplatting	12.167	23.428	0.727	0.327	27.773	26.194	0.783	0.310	24.421	25.967	0.805	0.278	21.256	25.186	0.788	0.284	21.838	23.603	0.756	0.315	21.255	25.483	0.831	0.237	25.474	25.106	0.785	0.266	30.305	26.757	0.835	0.323	13.332	25.591	0.756	0.333	21.806	25.399	0.752	0.309	26.107	25.201	0.735	0.341	28.610	23.984	0.813	0.248	44.293	25.959	0.879	0.304	11.910	23.919	0.778	0.309	44.749	26.985	0.841	0.244	29.897	29.787	0.925	0.164	14.122	28.450	0.879	0.177	32.048	25.091	0.760	0.324	29.102	24.051	0.726	0.323	26.158	24.670	0.814	0.244	25.792
FC-GS	12.167	19.383	0.669	0.200	54.895	21.244	0.711	0.211	32.313	21.725	0.740	0.201	30.955	19.918	0.704	0.197	33.698	19.506	0.677	0.235	47.107	21.018	0.764	0.145	47.620	20.245	0.700	0.180	59.600	21.360	0.783	0.226	34.130	20.410	0.672	0.287	27.717	22.185	0.744	0.187	51.984	21.944	0.679	0.214	59.713	18.734	0.675	0.202	77.679	20.569	0.815	0.221	38.369	19.574	0.635	0.251	98.962	22.455	0.761	0.163	41.207	22.249	0.805	0.117	20.750	22.031	0.770	0.136	41.766	21.437	0.724	0.189	57.524	20.038	0.668	0.205	57.524	19.908	0.718	0.179	48.940
Lightgaussian	13.667	24.320	0.764	0.292	392.138	26.259	0.774	0.340	175.644	26.575	0.819	0.276	264.895	25.488	0.792	0.291	224.205	24.352	0.778	0.310	281.697	26.659	0.850	0.228	361.130	26.216	0.819	0.236	407.525	26.586	0.833	0.331	111.166	25.979	0.750	0.347	165.273	26.066	0.773	0.299	246.503	25.464	0.728	0.351	194.879	26.239	0.887	0.162	654.902	26.044	0.879	0.309	119.238	26.508	0.874	0.195	800.511	27.936	0.863	0.236	282.481	30.644	0.935	0.157	165.865	30.311	0.910	0.143	412.095	26.265	0.799	0.294	385.430	25.039	0.770	0.287	335.331	25.648	0.844	0.221	347.001
gaussianspa	14.333	23.443	0.723	0.334	106.796	26.651	0.790	0.304	101.522	26.205	0.810	0.274	106.681	25.849	0.803	0.271	107.060	24.366	0.776	0.298	106.461	26.343	0.839	0.237	107.462	25.129	0.781	0.275	107.190	27.192	0.843	0.313	105.747	26.091	0.762	0.325	97.891	25.572	0.753	0.308	106.949	25.503	0.728	0.353	99.616	23.232	0.774	0.298	106.540	26.490	0.886	0.294	107.086	23.316	0.739	0.359	107.694	27.245	0.837	0.259	101.901	31.167	0.938	0.146	91.718	29.001	0.881	0.183	106.679	25.319	0.760	0.327	106.925	24.483	0.738	0.314	106.536	25.048	0.825	0.236	107.256
Octree-GS	14.333	23.048	0.706	0.339	79.979	26.623	0.793	0.292	80.002	26.631	0.831	0.239	114.507	21.659	0.652	0.427	46.653	20.328	0.603	0.465	21.491	21.168	0.654	0.440	13.764	21.287	0.584	0.453	15.707	24.323	0.757	0.410	7.028	23.993	0.683	0.407	33.737	23.689	0.661	0.401	30.328	24.550	0.656	0.417	34.681	20.510	0.625	0.393	26.897	25.199	0.846	0.348	13.338	25.085	0.838	0.245	145.552	22.835	0.619	0.481	8.832	21.762	0.746	0.399	2.186	25.606	0.773	0.301	39.751	22.981	0.653	0.416	32.689	20.824	0.542	0.496	10.875	22.034	0.688	0.390	86.941

Visual benchmark comparison across multiple PKU-GS scene categories and reconstruction methods. — Qualitative comparison of benchmark methods across representative PKU-GS scene categories.

Dataset Access

Download

Category-level release access for PKU-GS raw videos and processed results, presented as a primary section rather than a secondary resources card.

Dataset Downloads

Category-wise Access

Raw videos and processed outputs are both available now, and multipart processed releases are split into separate entries where needed.

Dataset download links by category, including raw videos and combined processed results.
Category	raw videos	Colmap + reconstruction
Standalone Buildings		partone parttwo
Architecture Complex
Sports Field
Natural Landscape
Sculpture
Waterfront Bridges