High-res multi-view results

This table lists the benchmark results for the high-res multi-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	high-res multi-view	indoor	outdoor	botani.	boulde.	bridge	door	exhibi.	lectur.	living.	lounge	observ.	old co.	statue	terrac.

MVELAS		82.73 1	82.35 1	83.86 1	87.90 2	67.65 1	95.45 1	92.18 1	85.06 1	72.58 1	92.27 1	52.46 1	95.78 2	84.18 1	79.06 2	88.16 1

ACMM		74.34 2	72.73 2	79.17 4	84.25 5	57.87 4	90.36 2	89.93 2	63.17 2	65.64 2	86.04 3	37.79 2	94.10 4	65.89 2	71.52 4	85.55 2

MSDG		69.54 3	65.73 4	80.96 2	85.11 4	63.71 2	84.19 6	83.28 7	56.95 5	54.50 5	84.03 4	31.50 6	96.03 1	42.87 4	69.13 5	83.14 3

LS3D		69.17 4	66.40 3	77.50 5	91.92 1	56.12 6	85.12 5	80.37 8	46.96 8	56.50 3	87.51 2	37.25 3	93.71 6	32.52 9	79.40 1	82.68 4

ACMH	63.93 1	68.62 5	64.81 5	80.03 3	85.22 3	63.25 3	87.72 4	84.66 4	60.83 3	51.57 7	79.43 8	27.29 7	94.96 3	41.76 5	64.84 8	81.88 5

CMPMVS	48.56 5	66.77 6	64.41 6	73.84 8	70.65 10	50.31 9	77.79 8	85.73 3	45.54 9	44.76 9	82.14 5	35.40 4	90.14 9	63.18 3	74.54 3	81.07 6
M. Jancosek, T. Pajdla: Multi-View Reconstruction Preserving Weakly-Supported Surfaces. CVPR 2011
OpenMVS_ROB	61.12 2	66.39 7	62.92 8	76.80 6	76.51 8	57.77 5	89.22 3	83.41 6	55.48 6	53.86 6	77.84 9	26.28 8	93.95 5	34.90 7	68.76 6	78.68 7

LTVRE_ROB	59.60 3	66.27 8	63.54 7	74.45 7	84.00 6	51.55 8	67.08 9	83.53 5	58.78 4	54.94 4	80.31 7	34.54 5	93.23 7	40.64 6	68.03 7	78.58 8
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
COLMAP_ROB	57.96 4	62.98 9	59.65 9	72.98 9	81.44 7	53.00 7	83.75 7	75.53 9	48.34 7	48.81 8	81.40 6	24.14 9	90.30 8	32.95 8	60.52 9	75.65 9
Johannes L. Schönberger, Enliang Zheng, Marc Pollefeys, Jan-Michael Frahm: Pixelwise View Selection for Unstructured Multi-View Stereo. ECCV 2016
Gipuma		34.91 10	31.44 10	45.30 11	70.99 9	39.64 10	19.85 12	72.56 10	20.10 12	16.16 11	21.47 13	5.08 11	71.16 10	13.07 11	43.70 11	25.08 13
S. Galliani, K. Lasinger, K. Schindler: Massively Parallel Multiview Stereopsis by Surface Normal Diffusion. ICCV 2015
PMVS	26.43 6	31.84 11	28.16 11	42.89 12	25.87 13	27.58 12	50.92 10	49.78 12	21.37 11	14.17 12	40.81 11	2.01 12	66.62 11	9.61 13	38.88 12	34.49 12
Y. Furukawa, J. Ponce: Accurate, dense, and robust multiview stereopsis. PAMI (2010)
MVE	24.84 7	24.35 12	19.77 12	38.09 13	34.56 12	26.92 13	26.57 11	38.87 13	11.73 13	11.37 13	27.44 12	1.37 13	50.42 13	11.41 12	14.60 13	36.93 11
Simon Fuhrmann, Fabian Langguth, Michael Goesele: MVE - A Multi-View Reconstruction Environment. EUROGRAPHICS Workshops on Graphics and Cultural Heritage (2014)
FPMVS				53.23 10	51.42 11	36.46 11		71.91 11	29.13 10	27.53 10	61.55 10	9.83 10	65.01 12	16.00 10	55.58 10	58.22 10

High-resolution multi-view benchmark