Optical flow evaluation results: R2.5 angle error

Optical flow evaluation results

Statistics: Average SD R2.5 R5.0 R10.0 A50 A75 A95
Error type: angle endpoint interpolation normalized interpolation

R2.5 angle error	avg.	Army (Hidden texture) GT im0 im1			Mequon (Hidden texture) GT im0 im1			Schefflera (Hidden texture) GT im0 im1			Wooden (Hidden texture) GT im0 im1			Grove (Synthetic) GT im0 im1			Urban (Synthetic) GT im0 im1			Yosemite (Synthetic) GT im0 im1			Teddy (Stereo) GT im0 im1
	rank	all	disc	untext	all	disc	untext	all	disc	untext	all	disc	untext	all	disc	untext	all	disc	untext	all	disc	untext	all	disc	untext
Classic+Area [31]	3.4	26.6 1	50.9 1	27.9 3	20.1 5	46.6 5	22.5 4	18.6 1	46.6 2	18.7 5	9.32 1	36.7 1	6.67 2	31.9 1	41.8 2	24.7 1	22.1 4	52.6 2	19.4 2	40.9 13	49.0 1	49.1 21	4.95 1	12.8 1	7.35 2
Adaptive [23]	6.5	27.0 2	52.6 2	19.8 1	21.9 6	47.8 7	22.6 6	20.5 5	47.8 7	19.7 6	10.3 3	39.9 3	6.31 1	45.6 24	52.6 23	51.3 22	17.4 1	48.2 1	13.9 1	34.8 9	56.9 9	19.8 11	6.04 2	15.2 2	7.54 3
TV-L1-improved [17]	8.2	27.7 3	57.9 3	20.5 2	18.2 3	44.9 3	19.1 3	19.4 3	47.7 6	17.0 4	10.1 2	38.5 2	6.75 3	35.9 4	46.0 6	27.3 2	43.7 24	70.2 22	47.5 25	51.4 21	60.5 12	50.3 22	10.3 6	26.8 8	10.8 8
Complementary OF [24]	9.3	51.9 16	74.9 16	59.3 17	14.2 1	41.6 2	13.7 1	20.4 4	46.6 2	19.7 6	22.2 15	40.8 4	21.0 17	36.0 5	43.4 3	38.5 9	33.9 13	63.8 11	31.6 11	31.1 7	51.9 4	36.2 14	18.9 15	34.4 17	29.4 14
Spatially variant [19]	9.4	41.1 7	74.5 15	43.0 8	22.3 7	45.9 4	25.7 10	19.2 2	46.7 4	16.2 1	13.7 6	43.2 7	8.89 6	38.6 11	47.1 10	40.1 12	30.4 9	65.2 13	32.3 12	28.1 4	55.4 6	18.2 8	28.0 21	40.0 21	38.0 22
Aniso. Huber-L1 [25]	9.5	33.9 4	65.1 5	32.8 4	34.0 17	54.0 15	40.0 17	27.9 17	55.0 13	38.4 18	15.2 11	49.9 12	12.0 11	35.3 3	44.6 5	28.5 4	23.9 5	55.9 4	20.7 3	50.6 19	62.1 14	39.7 15	8.15 5	20.7 4	8.39 4
Rannacher [27]	9.9	43.1 11	71.0 9	45.2 11	24.1 11	49.4 8	26.4 12	26.0 15	56.9 16	26.0 13	14.2 8	42.7 6	10.5 8	37.1 7	47.9 12	30.7 6	32.3 11	65.2 13	27.0 9	44.0 15	56.0 7	39.7 15	7.83 4	21.2 5	9.19 6
Brox et al. [5]	11.1	43.7 12	74.4 13	56.4 16	27.0 13	52.5 14	30.5 14	23.4 11	54.5 12	23.3 10	13.4 5	50.0 13	8.66 5	39.8 13	46.5 8	47.6 18	21.6 2	59.8 7	22.8 5	30.9 6	59.3 11	7.61 3	23.1 18	37.0 19	33.4 18
Second-order prior [8]	11.5	37.6 5	70.9 8	37.2 7	22.4 8	51.2 11	23.7 8	24.5 14	59.1 18	22.6 8	11.2 4	41.2 5	8.48 4	37.9 8	49.6 20	28.8 5	29.0 8	68.8 19	26.0 8	55.5 24	64.7 18	52.7 24	14.7 12	34.1 16	17.3 13
Occlusion bounds [26]	11.5	42.7 9	71.4 10	53.1 14	29.4 15	55.1 17	33.2 15	24.0 13	54.4 11	26.0 13	14.0 7	50.1 14	9.34 7	40.4 15	47.0 9	49.0 21	22.0 3	57.6 6	22.6 4	29.2 5	56.8 8	7.51 2	24.2 20	38.9 20	35.6 19
NL-TV-NCC [29]	11.5	46.1 13	68.3 6	43.4 9	25.2 12	55.3 18	23.5 7	21.8 8	47.2 5	16.9 3	16.9 13	44.1 9	11.9 10	38.5 10	49.1 18	27.3 2	36.5 17	65.7 15	34.9 13	46.2 16	75.7 28	45.7 19	12.6 11	28.8 10	9.10 5
F-TV-L1 [15]	11.6	66.8 23	84.2 21	77.3 22	27.1 14	52.0 13	29.1 13	27.2 16	57.3 17	24.2 11	24.1 16	52.0 16	19.5 15	39.3 12	47.7 11	39.3 10	24.2 6	56.5 5	24.9 6	33.2 8	53.7 5	20.1 12	6.69 3	18.6 3	5.94 1
CBF [12]	11.7	41.4 8	74.0 12	48.5 13	40.2 18	51.5 12	51.7 20	22.9 10	50.8 9	28.5 16	14.3 9	44.7 10	11.2 9	38.3 9	46.1 7	36.1 7	26.3 7	55.1 3	24.9 6	61.5 25	71.0 21	52.0 23	11.6 8	26.7 7	14.8 11
DPOF [18]	11.9	48.2 14	65.0 4	36.4 5	23.7 10	51.1 10	24.4 9	23.4 11	48.7 8	26.1 15	27.9 17	50.4 15	26.2 18	36.1 6	44.2 4	37.5 8	33.1 12	62.1 9	39.5 18	54.3 22	64.2 16	53.0 25	12.1 10	30.4 11	12.5 9
Dynamic MRF [7]	13.0	49.5 15	78.0 19	55.8 15	17.2 2	47.4 6	16.5 2	21.6 7	56.3 15	16.2 1	14.8 10	46.4 11	12.5 12	41.2 16	49.0 16	45.1 16	35.3 15	70.7 23	38.5 16	37.1 11	57.7 10	55.1 26	21.3 16	36.7 18	31.2 15
Multicue MRF [21]	13.7	59.3 18	77.1 18	68.2 19	22.5 9	41.2 1	26.0 11	22.1 9	43.7 1	25.7 12	33.4 19	43.3 8	32.3 19	35.2 2	39.7 1	41.6 13	40.3 21	61.0 8	40.0 19	66.4 28	71.2 22	74.5 30	18.8 14	28.4 9	33.0 17
Learning Flow [11]	14.6	42.9 10	70.7 7	44.8 10	30.2 16	54.7 16	34.7 16	28.2 18	55.9 14	32.3 17	17.6 14	57.1 18	12.6 13	44.0 21	52.6 23	47.8 19	30.5 10	64.6 12	30.3 10	46.9 17	63.3 15	42.8 18	14.7 12	31.6 12	16.3 12
Fusion [6]	15.8	40.5 6	75.6 17	45.9 12	20.0 4	50.0 9	22.5 4	20.8 6	52.8 10	22.8 9	16.2 12	52.8 17	13.5 14	43.1 18	49.0 16	47.5 17	39.6 20	67.8 18	43.8 23	63.9 27	75.0 26	46.3 20	35.5 25	42.6 23	53.3 27
SegOF [10]	17.2	56.3 17	71.9 11	37.1 6	57.3 24	62.9 22	68.3 24	46.0 22	69.0 21	57.2 25	41.0 21	59.5 19	37.2 21	43.5 19	48.3 14	56.4 26	38.2 18	69.6 21	39.1 17	17.9 1	64.5 17	3.40 1	22.7 17	33.0 14	32.0 16
Filter Flow [20]	19.4	62.9 19	74.4 13	60.8 18	42.8 19	60.1 20	49.4 19	42.5 20	66.0 19	51.2 19	52.1 25	69.5 22	50.2 24	44.8 23	49.7 21	54.4 24	41.9 23	66.7 16	43.6 22	74.3 30	88.9 30	42.6 17	10.6 7	21.6 6	12.6 10
SPSA-learn [13]	20.0	65.1 20	87.4 25	72.7 20	45.7 22	59.3 19	53.4 22	45.2 21	74.7 22	52.2 20	41.6 22	69.9 23	42.5 22	42.7 17	48.9 15	48.7 20	38.8 19	69.0 20	42.5 21	39.1 12	61.9 13	19.3 9	36.0 26	45.6 25	48.3 25
GraphCuts [14]	20.1	66.6 22	87.0 24	80.0 23	43.1 20	63.0 23	46.1 18	41.8 19	67.0 20	53.4 22	28.5 18	64.0 21	20.8 16	40.2 14	48.1 13	43.5 14	46.5 26	63.4 10	40.5 20	62.7 26	75.4 27	69.5 29	23.8 19	33.3 15	38.5 23
GroupFlow [9]	20.9	66.4 21	85.2 23	80.8 24	61.6 26	75.4 28	69.0 25	51.9 24	83.6 25	57.0 24	33.5 20	63.9 20	32.5 20	49.6 25	61.0 27	39.9 11	51.3 29	81.7 26	59.4 28	22.8 3	51.1 3	16.5 5	28.0 21	41.9 22	37.9 21
2D-CLG [1]	21.1	77.2 28	82.3 20	75.4 21	61.5 25	65.9 24	73.7 26	63.2 28	89.6 28	60.8 28	82.8 30	88.3 29	86.8 30	43.5 19	49.3 19	54.8 25	35.1 14	67.5 17	36.1 14	21.3 2	50.8 2	15.5 4	34.4 24	46.3 26	46.0 24
Black & Anandan [4]	21.2	70.3 25	88.0 26	84.1 26	45.5 21	61.4 21	52.0 21	47.4 23	77.3 23	52.9 21	42.3 23	77.5 25	42.8 23	44.0 21	51.8 22	45.0 15	35.9 16	75.9 24	38.3 15	50.8 20	71.3 23	17.8 7	29.8 23	42.7 24	37.6 20
Horn & Schunck [3]	24.6	74.1 27	93.2 28	86.9 27	49.1 23	73.8 26	53.9 23	53.1 25	89.0 27	54.6 23	50.9 24	81.4 27	52.4 25	51.3 26	58.8 25	54.3 23	41.2 22	82.3 27	44.6 24	55.0 23	74.3 25	19.6 10	40.7 27	56.8 27	48.8 26
FOLKI [16]	26.4	72.5 26	84.5 22	87.5 28	62.3 27	74.9 27	74.0 27	56.9 26	87.2 26	57.7 26	59.8 26	78.1 26	65.4 26	53.3 28	61.2 28	62.7 28	50.9 28	81.0 25	62.7 29	47.3 18	73.3 24	56.8 27	49.0 28	62.9 29	64.3 29
STOB [22]	26.5	67.5 24	90.3 27	82.1 25	72.2 29	84.7 31	84.8 29	58.4 27	94.0 30	58.1 27	78.1 28	82.5 28	84.6 29	55.4 29	61.5 29	68.1 30	49.4 27	83.7 29	57.7 27	36.2 10	68.1 20	26.2 13	50.3 29	60.0 28	65.2 30
Bipartite [30]	26.5	87.8 30	95.4 31	88.8 29	67.4 28	69.2 25	75.3 28	74.2 31	82.3 24	81.4 31	77.1 27	71.8 24	77.2 27	64.4 30	66.6 30	62.5 27	81.9 31	86.7 31	87.0 31	92.6 31	94.8 31	78.6 31	11.6 8	32.4 13	9.90 7
TI-DOFE [28]	27.3	90.7 31	94.6 30	97.1 31	76.9 31	79.5 30	89.4 31	73.1 30	96.1 31	74.4 29	84.6 31	93.6 31	88.3 31	52.9 27	59.9 26	63.9 29	44.5 25	83.6 28	53.5 26	42.9 14	68.0 19	17.0 6	50.5 30	65.5 30	62.4 28
Pyramid LK [2]	29.7	86.7 29	94.5 29	96.2 30	73.1 30	76.5 29	86.4 30	70.8 29	89.6 28	78.5 30	78.1 28	88.6 30	82.7 28	68.8 31	76.4 31	80.4 31	75.7 30	85.1 30	78.1 30	73.1 29	79.6 29	69.3 28	60.8 31	74.5 31	79.9 31

Move the mouse over the numbers in the table to see the corresponding images. Click to compare with the ground truth.

Color encoding
of flow vectors
Army - Ground-truth flow

References

Method	time*	frames	color	Reference and notes
[1] 2D-CLG	844	2	gray	The 2D-CLG method by Bruhn et al. as implemented by Stefan Roth. [A. Bruhn, J. Weickert, and C. Schnörr. Lucas/Kanade meets Horn/Schunck: combining local and global optic flow methods. IJCV 63(3), 2005.] Parameters were set to match the published performance on Yosemite sequence, which may not be optimal for other sequences.
[2] Pyramid LK	12	2	color	A modification of Bouguet's pyramidal implementation of Lucas-Kanade.
[3] Horn & Schunck	49	2	gray	A modern Matlab implementation of the Horn & Schunck method by Deqing Sun. Parameters set to optimize AAE on all training data.
[4] Black & Anandan	328	2	gray	A modern Matlab implementation of the Black & Anandan method by Deqing Sun.
[5] Brox et al.	18	2	color	T. Brox, A. Bruhn, N. Papenberg, and J. Weickert. High accuracy optical flow estimation based on a theory for warping. ECCV 2004. (Improved using separate robust functions as proposed in A. Bruhn and J. Weickert, Towards ultimate motion estimation, ICCV 2005; improved by training on the training set.)
[6] Fusion	2,666	2	color	V. Lempitsky, S. Roth, and C. Rother. Discrete-continuous optimization for optical flow estimation. CVPR 2008.
[7] Dynamic MRF	366	2	gray	B. Glocker, N. Paragios, N. Komodakis, G. Tziritas, and N. Navab. Optical flow estimation with uncertainties through dynamic MRFs. CVPR 2008. (Method improved since publication.)
[8] Second-order prior	14	2	gray	W. Trobin, T. Pock, D. Cremers, and H. Bischof. An unbiased second-order prior for high-accuracy motion estimation. DAGM 2008. (Method improved since publication; for details see W. Trobin, Ph.D. thesis, 2009.)
[9] GroupFlow	600	2	gray	X. Ren. Local Grouping for Optical Flow. CVPR 2008.
[10] SegOF	60	2	color	L. Xu, J. Chen, and J. Jia. Segmentation based variational model for accurate optical flow estimation. ECCV 2008.
[11] Learning Flow	825	2	gray	D. Sun, S. Roth, J.P. Lewis, and M. Black. Learning optical flow (SRF-LFC). ECCV 2008.
[12] CBF	69	2	color	W. Trobin, T. Pock, D. Cremers, and H. Bischof. Continuous energy minimization via repeated binary fusion. ECCV 2008. (Method improved since publication; for details see W. Trobin, Ph.D. thesis, 2009.)
[13] SPSA-learn	200	2	color	Y. Li and D. Huttenlocher. Learning for optical flow using stochastic optimization. ECCV 2008.
[14] GraphCuts	1,200	2	color	T. Cooke. Two applications of graph-cuts to image processing. DICTA 2008.
[15] F-TV-L1	8	2	gray	A. Wedel, T. Pock, J. Braun, U. Franke, and D. Cremers. Duality TV-L1 flow with fundamental matrix prior. IVCNZ 2008.
[16] FOLKI	1.4	2	gray	G. Le Besnerais and F. Champagnat. Dense optical flow by iterative local window registration. ICIP 2005.
[17] TV-L1-improved	2.9	2	gray	A. Wedel, T. Pock, C. Zach, H. Bischof, and D. Cremers. An improved algorithm for TV-L1 optical flow computation. Proceedings of the Dagstuhl Visual Motion Analysis Workshop 2008. Code at GPU4Vision.
[18] DPOF	261	2	color	C. Lei and Y.-H. Yang. Optical flow estimation on coarse-to-fine region-trees using discrete optimization. ICCV 2009.
[19] Spatially variant	2,100	2	color	Anonymous. Optical flow estimation with spatially-variant smoothness constraint. ICCV 2009 submission 1860.
[20] Filter Flow	34,000	2	color	S. Seitz and S. Baker. Filter flow. ICCV 2009.
[21] Multicue MRF	13,240	2	color	Anonymous. Optical flow estimation using discrete optimization with multi-cue weighted correlation and occlusion handling. ICCV 2009 submission 766.
[22] STOB	1,080	2	gray	Anonymous. Stochastic uncertainty models for motion estimation. ICCV 2009 submission 1000.
[23] Adaptive	9.2	2	gray	A. Wedel, D. Cremers, T. Pock, and H. Bischof. Structure- and motion-adaptive regularization for high accuracy optic flow. ICCV 2009.
[24] Complementary OF	44	2	color	H. Zimmer, A. Bruhn, J. Weickert, L. Valgaerts, A. Salgado, B. Rosenhahn, and H.-P. Seidel. Complementary optic flow. EMMCVPR 2009.
[25] Aniso. Huber-L1	2	2	gray	M. Werlberger, W. Trobin, T. Pock, A. Wedel, D. Cremers, and H. Bischof. Anisotropic Huber-L1 optical flow. BMVC 2009. Code at GPU4Vision.
[26] Occlusion bounds	300	3	color	Anonymous. Occlusion boundaries. NIPS 2009 submission 245.
[27] Rannacher	0.12	2	gray	J. Rannacher. Realtime 3D motion estimation on graphics hardware. Bachelor thesis, Heidelberg University, 2009.
[28] TI-DOFE	260	2	gray	C. Cassisa, S. Simoens, and V. Prinet. Two-frame optical flow formulation in an unwarped multiresolution scheme. CIARP 2009.
[29] NL-TV-NCC	20	2	color	Anonymous. Motion estimation with non-local total variation regularization. CVPR 2010 submission 778.
[30] Bipartite	120	2	gray	Anonymous. Dynamic bipartite matching. CVPR 2010 submission 69.
[31] Classic+Area	791	2	gray	Anonymous. Secrets of optical flow estimation and their principles. CVPR 2010 submission 477.

* The "time" column lists the reported runtime in seconds on the "Urban" sequence. Note that these runtimes are not normalized by processor speed or type.