EvalMVX: A Unified Benchmarking for Neural 3D Reconstruction under Diverse Multiview Setups

Zaiyan Yang; Jieji Ren; Xiangyi Wang; zonglin li; Xu Cao; Heng Guo; Zhanyu Ma; Boxin Shi

cs.CVFeb 27, 2026

EvalMVX: A Unified Benchmarking for Neural 3D Reconstruction under Diverse Multiview Setups

Authors: Zaiyan Yang, Jieji Ren, Xiangyi Wang, zonglin li, Xu Cao, Heng Guo, Zhanyu Ma, Boxin Shi

Organizations: Beijing University of Posts and Telecommunications, Beijing, China · School of Mechanical Engineering, Shanghai Jiao Tong University, Shanghai, China · Independent Researcher · Xiong’an Aerospace Information Research Institute, Xiong’an, China ·

Abstract

Recent advancements in neural surface reconstruction have significantly enhanced 3D reconstruction. However, current real world datasets mainly focus on benchmarking multiview stereo (MVS) based on RGB inputs. Multiview photometric stereo (MVPS) and multiview shape from polarization (MVSfP), though indispensable on high-fidelity surface reconstruction and sparse inputs, have not been quantitatively assessed together with MVS. To determine the working range of different MVX (MVS, MVSfP, and MVPS) techniques, we propose EvalMVX, a real-world dataset containing $25$ objects, each captured with a polarized camera under $20$ varying views and $17$ light conditions including OLAT and natural illumination, leading to $8,500$ images. Each object includes aligned ground-truth 3D mesh, facilitating quantitative benchmarking of MVX methods simultaneously. Based on our EvalMVX, we evaluate $13$ MVX methods published in recent years, record the best-performing methods, and identify open problems under diverse geometric details and reflectance types. We hope EvalMVX and the benchmarking results can inspire future research on multiview 3D reconstruction.

Figures & tables

Figure 1: Dataset overview. EvalMVX contains 25 objects (left), covering diverse shapes and materials. From top to bottom, the materials range from simple diffuse to complex metallic and reflective ones, while the complexity of object shapes increases from left to right. With aligned GT shapes (middle), we can quantitatively evaluate MVX methods on the same object (right) and reveal their performance on detailed shape recovery and robustness against complex material.

Dataset	EvalMVS	EvalMVPS	EvalMVSfP	3D GT	#Objects	#Lights	#Views
DTU [ 18 ]	✓	$\times$	$\times$	Point cloud	80	1/7 Env. a	49 $\sim$ 64
Mobilebrick [ 24 ]	✓	$\times$	$\times$	Mesh	153	1 Env.	30 $\sim$ 134
Stanford-ORB [ 22 ]	✓	$\times$	$\times$	Mesh	14	3/7 Env.	70
Objects with Lighting [ 36 ]	✓	$\times$	$\times$	None	8	3/72 Env.	42 $\sim$ 67
CO3D [ 33 ]	✓	$\times$	$\times$	Point cloud	19K	Env. (Case dependent)	Case dependent
MVImageNet [ 46 ]	✓	$\times$	$\times$	None	220K	Env. (Case dependent)	Case dependent

Table 1: Summary of MVX datasets and their contents.

Figure 2: Capture setup and data processing of EvalMVX. ( Top ) Our capture system contains a polarization camera and 16 evenly distributed LEDs for recording multiview and multi-light images. ( Bottom ) Polarization information can be extracted from the snapshot polarization image.

Method	MVS				MVPS				MVSfP
Method	NeuS2 [ 40 ]	NeRO [ 29 ]	PET-NeuS [ 41 ]	GaussianSurfels [ 9 ]	RnbNeuS [ 3 ]	SuperNormal [ 6 ]	VMINer [ 11 ]	WildLight [ 8 ]	MVAS [ 4 ]	NeRSP [ 14 ]	PISR [ 7 ]
Face	0.234	0.411	0.303	0.540	0.614	0.277	0.400	0.250	0.389	0.330	0.308
Pig	0.312	0.282	0.408	0.720	0.379	0.256	0.470	0.265	0.446	0.447	0.286
Owl1	0.327	0.306	0.360	0.636	0.443	0.288	0.319	0.342	0.354	0.371	0.423
Monkey	0.487	0.345	0.664	0.646	0.393	0.325	0.530	0.335	0.365	0.501	0.311
Owl2	0.383	0.321	0.393	0.639	0.504	0.300	0.415	0.377	0.444	0.405	0.390

Table 2: Benchmark results on EvalMVX measured by CD in millimeter over 25 objects, where the smallest and second smallest CD values are shown in bold and underlined, respectively.

Figure 3: Inaccurate surface normal inputs from Uni-MS-PS [ 15 ] due to complex reflectance influence the shape recovery from SuperNormal [ 6 ] .

Figure 4: The difference between the real-captured AoP map and azimuth map due to sensor noise degrades the performance of MVSfP methods.

Figure 5: Best-performing method for each object in EvalMVX, with CD displayed in brackets. Summary results shown in left and right correspond to using 20 and 10 input views.

Figure 6: Visualization of accuracy and efficiency across different methods.

Figure 7: SuperNormal [ 6 ] with integrated depth improves reconstructions considering per-view shape continuity.

Figure 8: ( Left ) Distribution of camera poses. ( Right ) Relative positions between the LEDs and camera.

Object	Resolution	Per-view Optimization Time	Average IoU
Pig	153×128 $\rightarrow$ 1224×1024	6s $\rightarrow$ 2min	0.9910 $\rightarrow$ 0.9917
Cambodia	153×128 $\rightarrow$ 1224×1024	30s $\rightarrow$ 20min	0.9882 $\rightarrow$ 0.9891

Table 3: Comparison of mesh alignment under different resolutions.

Figure 9: ( Left ) We align scanned mesh to camera views guided by GT masks. ( Right ) Curvature distribution of EvalMVX calculated from GT normals quantifies the shape complexity of each object.

Method	MVS				MVPS				MVSfP
Method	NeuS2 [ 40 ]	NeRO [ 29 ]	PET-NeuS [ 41 ]	GaussianSurfels [ 9 ]	RnbNeuS [ 3 ]	SuperNormal [ 6 ]	VMINer [ 11 ]	WildLight [ 8 ]	MVAS [ 4 ]	NeRSP [ 14 ]	PISR [ 7 ]
Face	0.254	0.281	0.303	0.670	0.541	0.260	0.370	0.366	0.420	0.588	0.396
Pig	0.324	0.279	0.461	0.705	0.386	0.254	0.456	0.360	0.498	0.424	0.355
Owl1	0.352	0.335	0.423	0.759	0.424	0.312	0.541	0.532	0.419	0.424	0.439
Monkey	0.561	0.346	0.698	0.736	0.425	0.308	0.457	0.441	0.428	0.514	0.337
Owl2	0.391	0.410	0.478	0.725	0.503	0.315	0.614	0.513	0.552	0.490	0.444

Table 4: Benchmark results on EvalMVX with 10 views - CD evaluation in millimeter over all 25 objects, where the smallest and second smallest ones are shown in bold and underlined, respectively.

Method	MVS				MVPS				MVSfP
Method	NeuS2 [ 40 ]	NeRO [ 29 ]	PET-NeuS [ 41 ]	GaussianSurfels [ 9 ]	RnbNeuS [ 3 ]	SuperNormal [ 6 ]	VMINer [ 11 ]	WildLight [ 8 ]	MVAS [ 4 ]	NeRSP [ 14 ]	PISR [ 7 ]
Face	5.96	3.36	4.88	12.32	4.19	2.22	6.92	3.95	8.09	7.52	14.98
Pig	9.57	4.58	8.84	16.53	5.62	4.15	9.40	5.06	8.86	9.60	8.69
Owl1	7.71	5.17	12.84	12.90	5.46	3.35	7.28	7.00	6.58	7.51	9.51
Monkey	16.64	5.62	25.82	18.51	6.20	4.41	17.89	7.08	8.42	13.05	7.60
Owl2	12.35	6.36	11.90	17.73	7.24	3.98	10.64	9.16	9.97	10.28	11.15

Table 5: Benchmark results on EvalMVX- MAE evaluation in degree over all 25 objects, where the smallest and second smallest ones are shown in bold and underlined, respectively.

Method	MVS						MVPS				MVSfP
Method	NeuS [ 39 ]	NeDUF [ 28 ]	NeuS2 [ 40 ]	NeRO [ 29 ]	PET-NeuS [ 41 ]	GaussianSurfels [ 9 ]	RnbNeuS [ 3 ]	SuperNormal [ 6 ]	VMINer [ 11 ]	WildLight [ 8 ]	MVAS [ 4 ]	NeRSP [ 14 ]	PISR [ 7 ]
Face	0.314	0.267	0.234	0.411	0.303	0.540	0.614	0.277	0.400	0.250	0.389	0.330	0.308
Pig	0.292	0.343	0.312	0.282	0.408	0.720	0.379	0.256	0.470	0.265	0.446	0.447	0.286
Owl1	0.306	0.413	0.327	0.306	0.360	0.636	0.443	0.288	0.319	0.342	0.354	0.371	0.423
Monkey	0.548	0.862	0.487	0.345	0.664	0.646	0.393	0.325	0.530	0.335	0.365	0.501	0.311
Owl2	0.319	0.392	0.383	0.321	0.393	0.639	0.504	0.300	0.415	0.377	0.444	0.405	0.390

Table 6: Extended benchmark results on EvalMVX, including the additional NeuS [ 39 ] and NeDUF [ 28 ] results.

Figure 10: Recovered shapes and error maps of Face .

Figure 11: Recovered shapes and error maps of Pig .

Figure 12: Recovered shapes and error maps of Owl1 .

Figure 13: Recovered shapes and error maps of Monkey .

Figure 14: Recovered shapes and error maps of Owl2 .

Figure 15: Recovered shapes and error maps of Bell .

Figure 16: Recovered shapes and error maps of Duck .

Figure 17: Recovered shapes and error maps of Monk1 .

Figure 18: Recovered shapes and error maps of Tiger .

Figure 19: Recovered shapes and error maps of SilverDog .

Figure 20: Recovered shapes and error maps of Snail .

Figure 21: Recovered shapes and error maps of Swan .

Figure 22: Recovered shapes and error maps of Yoga .

Figure 23: Recovered shapes and error maps of Hugo .

Figure 24: Recovered shapes and error maps of Chess .

Figure 25: Recovered shapes and error maps of BlackDog .

Figure 26: Recovered shapes and error maps of Monk2 .

Figure 27: Recovered shapes and error maps of Bird .

Figure 28: Recovered shapes and error maps of Caw .

Figure 29: Recovered shapes and error maps of Knight .

Figure 30: Recovered shapes and error maps of Frog .

Figure 31: Recovered shapes and error maps of Sphinx .

Figure 32: Recovered shapes and error maps of Cambodia .

Figure 33: Recovered shapes and error maps of Mummy .

Figure 34: Recovered shapes and error maps of Dragon .

3D Reconstruction Benchmarks

See all 15 papers

Can These Views Be One Scene? Evaluating Multiview 3D Consistency when 3D Foundation Models Hallucinate
May 18, 2026Soumava Paul, Prakhar Kaushik, Alan Yuille3D Foundation Models Multi-View Consistency
CCVL, Johns Hopkins University
3DReflecNet: A Large-Scale Dataset for 3D Reconstruction of Reflective, Transparent, and Low-Texture Objects
May 11, 2026Zhicheng Liang, Haoyi Yu, Boyan Li +63D Reconstruction Benchmarks 3D Reconstruction
The Chinese University of Hong Kong, Shenzhen · Capital Normal University · University of Southern California
A Comparison of Multi-View Stereo Methods for Photogrammetric 3D Reconstruction: From Traditional to Learning-Based Approaches
Apr 11, 2026Yawen Li, George Vosselman, Francesco NexAerial 3D Reconstruction 3D Reconstruction
Faculty of Geo-Information Science and Earth Observation (ITC), University of Twente, Enschede, The Netherlands

May 18, 2026J/K move · Enter open · S save

Can These Views Be One Scene? Evaluating Multiview 3D Consistency when 3D Foundation Models Hallucinate

Soumava Paul, Prakhar Kaushik, Alan Yuille

CCVL, Johns Hopkins University

3D Foundation Models Multi-View Consistency Multi-View Geometry 3D Reconstruction Benchmarks 3D Reconstruction 3D Representation Learning

Open paper AskPDF arXiv

Multiview 3D evaluation assumes that the images being scored are observations of one static 3D scene. This assumption can fail in NVS and sparse-view reconstruction: inputs or generated outputs may contain artifacts, outlier frames, repeated views, or noise, yet still receive high 3D consistency scores. Existing reference-based metrics require ground truth, while ground-truth-free metrics such as MEt3R depend on learned reconstruction backbones whose failure modes are poorly characterized. We study this reliability problem by comparing neural reconstruction priors with classical geometric verification. We introduce \benchmark, a controlled robustness benchmark for multiview 3D consistency, and a parametric family that decomposes neural metrics into backbone, residual, and aggregation components. This family recovers MEt3R and yields variants up to

3\times

more robust. Our analysis shows that VGGT, MASt3R, DUSt3R, and Fast3R can hallucinate dense geometry and cross-view support for unrelated scenes, repeated images, and random noise. We introduce COLMAP-based metrics that use matches, registration, dense support, and reconstruction failure as failure-aware consistency signals. On real NVS outputs and a structured human study, these metrics achieve up to

4\times

higher correlation with human judgments than MEt3R.

Abstract

3D Reconstruction Benchmarks

Explore similar work

Can These Views Be One Scene? Evaluating Multiview 3D Consistency when 3D Foundation Models Hallucinate

3DReflecNet: A Large-Scale Dataset for 3D Reconstruction of Reflective, Transparent, and Low-Texture Objects

A Comparison of Multi-View Stereo Methods for Photogrammetric 3D Reconstruction: From Traditional to Learning-Based Approaches