Low-res many-view results

This table lists the benchmark results for the low-res many-view scenario. The following metrics are evaluated:

Accuracy [%]: The fraction of the reconstruction which is closer to the ground truth than the evaluation threshold distance (*). Larger is better.
Completeness [%]: The fraction of the ground truth which is closer to the reconstruction than the evaluation threshold distance (*). Larger is better.
F₁ score [%]: The harmonic mean of accuracy and completeness, used to rank methods based on both metrics. Larger is better.
Time [s]: The runtime of the method. Smaller is better.

(*) For exact definitions, detailing how potentially incomplete ground truth is taken into account, see our paper.

The datasets are grouped into different categories, and result averages are computed for a category and method if results of the method are available for all datasets within the category. Note that the category "all" includes both the high-res multi-view and the low-res many-view scenarios.

Methods with suffix _ROB may participate in the Robust Vision Challenge.

Click a dataset result cell to show a visualization of the reconstruction. For training datasets, ground truth and accuracy / completeness visualizations are also available. The visualizations may not work with mobile browsers.

Method	all	low-res many-view	indoor	outdoor	lakeside	sand box	storage room	storage room 2	tunnel

DeepPCF-MVS	96.37 2	92.61 1	89.11 1	94.94 1	95.58 1	95.07 1	85.62 1	92.60 1	94.17 1

DeepC-MVS_fast	96.70 1	92.27 2	88.75 2	94.63 2	95.15 3	94.80 2	85.14 3	92.35 2	93.92 2
Andreas Kuhn, Christian Sormann, Mattia Rossi, Oliver Erdler, Friedrich Fraundorfer: DeepC-MVS: Deep Confidence Prediction for Multi-View Stereo Reconstruction. 3DV 2020
DeepC-MVS	95.98 3	91.43 3	87.48 4	94.06 4	94.52 5	94.15 5	83.69 6	91.26 3	93.50 3
Andreas Kuhn, Christian Sormann, Mattia Rossi, Oliver Erdler, Friedrich Fraundorfer: DeepC-MVS: Deep Confidence Prediction for Multi-View Stereo Reconstruction. 3DV 2020
TAPA-MVS	93.98 7	90.68 4	88.05 3	92.44 7	92.42 10	93.52 8	85.24 2	90.85 6	91.36 14
Andrea Romanoni, Matteo Matteucci: TAPA-MVS: Textureless-Aware PAtchMatch Multi-View Stereo. ICCV 2019
TAPA-MVS(SR)		90.57 5	87.48 4	92.63 6	92.64 8	93.17 11	83.91 5	91.04 4	92.10 12

PCF-MVS	93.45 11	90.50 6	85.08 14	94.11 3	95.54 2	93.44 10	82.27 12	87.88 16	93.34 4
Andreas Kuhn, Shan Lin, Oliver Erdler: Plane Completion and Filtering for Multi-View Stereo Reconstruction. GCPR 2019
COLMAP(SR)		90.17 7	85.79 12	93.08 5	92.63 9	93.87 6	82.51 11	89.07 14	92.74 6

PLC	95.07 4	89.96 8	87.00 6	91.94 12	90.65 19	92.68 17	83.01 8	90.98 5	92.48 7
Jie Liao, Yanping Fu, Qingan Yan, Chunxia xiao: Pyramid Multi-View Stereo with Local Consistency. Pacific Graphics 2019
COLMAP(base)		89.91 9	86.37 9	92.28 9	91.56 15	92.95 14	82.53 10	90.21 8	92.33 8

ACMH+	92.99 14	89.84 10	85.96 11	92.43 8	92.28 11	92.84 16	81.89 13	90.03 9	92.17 10

GSE		89.80 11	86.71 7	91.87 13	90.87 18	92.62 18	83.49 7	89.92 10	92.11 11

ACMM	93.85 9	89.42 12	85.42 13	92.10 11	91.09 17	92.98 13	81.16 15	89.68 11	92.23 9
Qingshan Xu and Wenbing Tao: Multi-Scale Geometric Consistency Guided Multi-View Stereo. CVPR 2019
ACMP	93.49 10	89.04 13	84.30 17	92.21 10	90.34 20	93.47 9	79.54 19	89.07 14	92.81 5
Qingshan Xu and Wenbing Tao: Planar Prior Assisted PatchMatch Multi-View Stereo. AAAI 2020
LTVRE_ROB	92.95 15	89.00 14	86.32 10	90.78 19	91.67 14	89.54 26	84.87 4	87.77 17	91.13 16
Andreas Kuhn, Heiko Hirschmüller, Daniel Scharstein, Helmut Mayer: A TV Prior for High-Quality Scalable Multi-View Stereo Reconstruction. International Journal of Computer Vision 2016
ACMH	92.88 16	88.89 15	84.55 15	91.78 14	91.37 16	92.41 19	79.88 17	89.21 12	91.56 13
Qingshan Xu and Wenbing Tao: Multi-Scale Geometric Consistency Guided Multi-View Stereo. CVPR 2019
IB-MVS	91.98 17	88.88 16	86.61 8	90.39 20	92.23 12	92.94 15	82.93 9	90.30 7	86.00 24
Christian Sormann, Mattia Rossi, Andreas Kuhn and Friedrich Fraundorfer: IB-MVS: An Iterative Algorithm for Deep Multi-View Stereo based on Binary Decisions. BMVC 2021
COLMAP_ROB	93.27 12	88.32 17	84.41 16	90.93 18	89.38 22	92.11 20	79.72 18	89.10 13	91.28 15
Johannes L. Schönberger, Enliang Zheng, Marc Pollefeys, Jan-Michael Frahm: Pixelwise View Selection for Unstructured Multi-View Stereo. ECCV 2016
3Dnovator	94.51 5	87.43 18	80.90 23	91.78 14	94.76 4	94.22 3	78.96 20	82.85 23	86.35 21

HY-MVS	93.96 8	87.12 19	82.85 20	89.96 22	89.33 23	93.07 12	78.79 21	86.90 19	87.48 19

3Dnovator+	94.38 6	87.07 20	80.26 24	91.61 16	94.34 6	94.17 4	78.01 22	82.52 24	86.34 22

A-TVSNet + Gipuma		86.96 21	82.91 19	89.65 23	88.10 25	91.71 21	80.23 16	85.60 22	89.15 18

BP-MVSNet		86.71 22	81.68 21	90.06 21	91.85 13	87.55 31	75.75 25	87.61 18	90.77 17
Christian Sormann, Patrick Knöbelreiter, Andreas Kuhn, Mattia Rossi, Thomas Pock, Friedrich Fraundorfer: BP-MVSNet: Belief-Propagation-Layers for Multi-View-Stereo. 3DV 2020
OpenMVS	93.04 13	86.34 23	79.08 25	91.18 17	93.50 7	93.77 7	76.56 23	81.60 27	86.26 23

LPCS		85.52 24	83.61 18	86.79 25	84.99 31	88.49 28	81.34 14	85.87 21	86.89 20

PVSNet_0	88.72 19	82.59 25	75.14 29	87.55 24	88.79 24	88.38 29	73.35 28	76.93 31	85.48 25

PVSNet	91.96 18	81.42 26	74.82 30	85.83 28	87.90 26	86.77 32	74.82 26	74.81 35	82.83 26

R-MVSNet		81.26 27	77.44 26	83.81 30	82.82 32	86.25 33	74.52 27	80.36 29	82.37 27

CIDER		80.48 28	72.14 32	86.05 26	87.66 27	89.77 24	71.75 29	72.53 39	80.72 29
Qingshan Xu and Wenbing Tao: Learning Inverse Depth Regression for Multi-View Stereo with Correlation Cost Volume. AAAI 2020
test_1205		80.31 29	75.46 28	83.55 31	86.48 29	90.90 22	70.38 31	80.54 28	73.26 41

test_1126		80.30 30	71.93 34	85.88 27	89.75 21	90.41 23	70.18 32	73.68 37	77.50 33

OpenMVS_ROB	86.42 20	80.25 31	71.88 35	85.82 29	87.52 28	89.56 25	70.53 30	73.24 38	80.39 30

ANet-0.75		77.73 32	74.78 31	79.70 34	77.01 39	84.71 36	67.34 36	82.21 26	77.39 34

P-MVSNet		77.69 33	81.61 22	75.07 42	71.91 46	77.33 44	76.51 24	86.71 20	75.98 37

CPR_FA		74.23 34	72.06 33	75.68 40	74.30 44	70.69 50	68.60 34	75.52 33	82.05 28

ANet		74.07 35	68.26 37	77.94 38	77.01 39	84.71 36	64.75 37	71.76 40	72.09 42

AttMVS		72.35 36	68.13 38	75.16 41	70.98 50	80.63 40	68.09 35	68.16 42	73.87 40

mvs_zhu_1030		72.30 37	62.31 42	78.96 36	77.79 38	83.27 38	55.37 43	69.25 41	75.83 38

Pnet-new-		70.55 38	53.66 56	81.81 33	81.27 34	85.23 35	51.15 52	56.17 64	78.92 32

CasMVSNet(SR_A)		69.66 39	55.07 52	79.38 35	80.08 35	87.62 30	46.56 58	63.58 54	70.44 44

Pnet_fast		68.89 40	47.01 67	83.47 32	81.84 33	89.30 27	38.61 69	55.42 65	79.28 31

unsupervisedMVS_cas		68.86 41	58.51 47	75.77 39	74.54 43	80.76 39	51.51 50	65.51 45	72.00 43

CasMVSNet(base)		68.63 42	53.70 55	78.58 37	79.92 36	85.96 34	44.83 60	62.57 56	69.87 45

A1Net		68.49 43	75.64 27	63.72 52	63.53 57	51.39 70	68.88 33	82.41 25	76.23 36

MVSNet_plusplus		64.09 44	48.06 65	74.77 43	85.29 30	62.00 55	31.11 73	65.01 48	77.03 35

Pnet-blend		64.02 45	52.13 60	71.95 44	75.81 41	78.00 42	40.66 65	63.61 52	62.05 52

Pnet-blend++		64.02 45	52.13 60	71.95 44	75.81 41	78.00 42	40.66 65	63.61 52	62.05 52

unMVSv1		63.91 47	62.59 41	64.80 51	66.39 53	68.00 53	61.62 38	63.55 55	60.01 57

MVSNet		63.58 48	56.25 49	68.47 50	66.00 54	71.12 47	48.36 55	64.13 49	68.29 47

MVSCRF		63.13 49	52.16 59	70.45 46	70.18 51	75.83 46	51.17 51	53.16 69	65.33 48

RMVSNet		62.69 50	69.01 36	58.48 60	69.15 52	58.56 59	59.72 39	78.30 30	47.75 67

metmvs_fine		60.95 51	66.50 39	57.25 62	64.00 55	51.92 66	59.04 40	73.97 36	55.84 62

Snet		60.77 52	47.18 66	69.82 47	77.90 37	61.73 56	39.61 68	54.76 66	69.82 46

Pnet-eth		60.43 53	65.87 40	56.80 63	63.59 56	45.39 74	55.58 42	76.16 32	61.43 56

MVSNet_++		59.91 54	45.04 69	69.82 47	74.28 45	60.08 57	14.97 76	75.11 34	75.10 39

DPSNet		59.89 55	54.16 53	63.70 53	71.39 47	70.87 48	52.09 48	56.24 62	48.85 64

hgnet		59.89 55	54.16 53	63.70 53	71.39 47	70.87 48	52.09 48	56.24 62	48.85 64

MVE	62.14 22	59.41 57	60.38 43	58.77 59	50.39 72	69.00 52	56.98 41	63.78 51	56.91 61
Simon Fuhrmann, Fabian Langguth, Michael Goesele: MVE - A Multi-View Reconstruction Environment. EUROGRAPHICS Workshops on Graphics and Cultural Heritage (2014)
test_1124		58.90 58	44.17 71	68.72 49	62.60 62	78.44 41	38.12 70	50.22 70	65.10 49

test3		57.95 59	60.13 44	56.50 65	54.75 69	51.79 67	53.99 44	66.26 43	62.98 50

PSD-MVSNet		57.63 60	58.95 45	56.74 64	56.38 63	52.03 64	52.75 45	65.16 47	61.82 55

SGNet		57.44 61	58.95 45	56.44 66	55.09 66	52.01 65	52.63 46	65.27 46	62.22 51

MVSNet + Gipuma		57.07 62	53.08 57	59.73 58	62.91 60	57.42 61	48.41 54	57.74 60	58.88 58

CasMVSNet(SR_B)		56.87 63	55.95 50	57.48 61	56.06 64	67.85 54	47.86 57	64.05 50	48.53 66

TVSNet		56.87 63	57.97 48	56.13 67	55.48 65	50.99 71	49.98 53	65.96 44	61.93 54

F/T MVSNet+Gipuma		55.78 65	49.83 64	59.76 57	63.10 58	57.51 60	46.34 59	53.31 68	58.66 59

example		55.06 66	44.75 70	61.93 55	71.21 49	69.85 51	42.82 62	46.68 71	44.73 71

firsttry		54.62 67	52.74 58	55.88 68	54.90 67	55.69 62	48.02 56	57.46 61	57.06 60

QQQNet		52.14 68	55.46 51	49.93 71	54.81 68	51.46 69	52.58 47	58.34 57	43.53 72

PMVS	61.03 23	49.24 69	32.69 76	60.27 56	62.70 61	76.26 45	42.99 61	22.38 76	41.86 74
Y. Furukawa, J. Ponce: Accurate, dense, and robust multiview stereopsis. PAMI (2010)
CCVNet		49.22 70	46.91 68	50.75 70	53.16 71	55.57 63	40.21 67	53.61 67	43.53 72

unMVSmet		48.01 71	40.73 73	52.86 69	53.90 70	58.72 58	36.72 72	44.74 73	45.96 68

ternet		47.76 72	49.94 62	46.31 73	48.51 73	45.06 75	41.53 63	58.34 57	45.36 69

SVVNet		47.76 72	49.94 62	46.31 73	48.51 73	45.06 75	41.53 63	58.34 57	45.36 69

confMetMVS		44.56 74	41.82 72	46.39 72	46.63 75	51.78 68	37.38 71	46.25 72	40.77 75

test_1120		40.25 75	35.81 75	43.21 75	30.29 77	50.38 72	29.89 75	41.72 75	48.95 63

Cas-MVS_preliminary		38.91 76	36.73 74	40.37 76	37.57 76	48.76 73	30.59 74	42.87 74	34.76 76

CMPMVS	66.06 21	9.32 77	0.44 78	15.23 77	6.17 78	39.51 77	0.89 78	0.00 78	0.00 78
M. Jancosek, T. Pajdla: Multi-View Reconstruction Preserving Weakly-Supported Surfaces. CVPR 2011
FADENet		4.02 78	4.30 77	3.83 78	6.03 79	3.71 78	5.02 77	3.58 77	1.74 77

dnet		0.00 79	0.00 79	0.00 79	0.00 80	0.00 79	0.00 79	0.00 78	0.00 78

UnsupFinetunedMVSNet					63.10 58

Low-resolution many-view benchmark