What’s new
21.03.20 | We published evaluation results for FBA Matting [22]. |
30.01.20 | We published evaluation results for Matting with Background Estimation [21]. |
08.12.17 | We published evaluation results for Information-Flow Matting [20]. |
07.04.17 | Subjective study results are now available. |
26.12.16 | We published evaluation results for Self-Adaptive Matting [17]. |
16.11.16 | We published evaluation results for Deep Matting [16]. |
6.04.16 | We published evaluation results for Sparse Sampling Matting [14]. |
15.12.15 | 1) New sequences with natural hairs including 3 public sequences. 2) New metrics of temporal coherency chosen by careful analysis (see [3]). 3) New trimap generation method (more natural-looking and accurate). 4) Better ground-truth quality owing to correction of lightning changes during capturing. 5) Improvement in website loading speed and interface. |
7.09.15 | We published the paper with our benchmark description [3]. |
30.12.14 | We published results for multiple levels of trimaps; use drop-down menu at the top left corner to switch levels. |
29.12.14 | We added general ranking to the rating table. |
10.11.14 | “Sparse codes as Alpha Matte” was added. |
26.09.14 | Source sequences are available for online view now. Full screen mode was added. |
30.08.14 | Composite sequences are available now. |
27.08.14 | “Refine edge tool in Adobe After Effects” was added. |
25.08.14 | The official opening. |
Overview
Introduction
The VideoMatting project is the first public objective benchmark for video-matting methods. It contains scatter plots and rating tables for different quality metrics. In addition, results for participating methods are available for viewing on a player equipped with a movable zoom region. We believe our work will help rank existing methods and aid developers of new methods in improving their results.
Datasets
The data set consists of five moving objects captured in front of a green plate and seven captured using the stop-motion procedure described below. We composed the objects over a set of background videos with various levels of 3D camera motion, color balance, and noise. We published ground-truth data for two stop-motion sequences and hid the rest to ensure fairness of the comparison.
Using thresholding and morphological operations on ground-truth alpha mattes, we generated narrow trimaps. Then, we dilated the results using graphcut-based energy minimization which provides us with more handmade-looking trimaps than common morphological dilation.
Chroma Keying
Chroma keying is a common practice of the cinema industry: the cinematographer captures an actor in front of a green or blue screen, then the VFX expert replaces the background using special software. Our evaluation uses five green-screen video sequences with a significant amount of semitransparency (e.g., hair or motion blur), provided to us by Hollywood camera work. We extract alpha mattes and corresponding foregrounds using The Foundry Keylight. Chroma keying enables us to get alpha mattes of natural-looking objects with arbitrary motion. Nevertheless, this technique can’t guarantee that the alpha maps are natural, because it assumes the screen color is absent from the foreground object. To get alpha maps that have a very natural appearance, we use the stop-motion method.
Stop Motion
We designed the following procedure to perform stop-motion capture: A fuzzy toy is placed on the platform in front of an LCD monitor. The toy rotates in small, discrete steps along a predefined 3D trajectory, controlled by two servos connected to a computer. After each step the digital camera in front of the setup captures the motionless toy against a set of background images. At the end of this process, the toy is removed and the camera again captures all of the background images.
We paid special attention to avoiding reflections of the background screen in the foreground object. These reflections can lead to false transparency that is especially noticeable in nontransparent regions. To reduce the amount of reflection we used checkerboard background images instead of solid colors, thereby adjusting the mean color of the screen to be the same for each background.
At the end we corrected global lighting changes caused by light bulb flickering. Thus finally we obtain alpha mattes with less than 1% of noise level. The detailed description of ground-truth extraction methods is given in [3].
Evaluation Methodology
Our comparison includes both image- and video-matting methods. We apply each matting method to the videos in our data set, and then compare the results using the following metrics of per-pixel accuracy and temporal coherency (look into our paper [3] for comparison of different metrics):
Here denotes total number of pixels, and denote transparency values of video matting under consideration and ground truth correspondingly at pixel of frame , and denotes motion vector at pixel . We use optical-flow algorithm [11] computed for ground-truth sequences. It is worth noting that motion-aware metrics will not give unfair advantage to matting methods based on the similar motion estimation method since they do not have ground truth sequence. The detailed description of used quality metrics is given in [3].
Public Sequences
For the training purposes we publish here three test sequences with their ground-truth transparency maps. Developers and researchers are welcome to use these sequences, but we ask to cite us [3]
Participate
We invite developers of video-matting methods to use our benchmark. We will evaluate the submitted data and report scores to the developer. In cases where the developer specifically grants permission, we will publish the results on our site. We can also publish anonymous scores for blind-reviewed papers. To participate, simply follow these steps:
- Download the data set containing our sequences: City, Flowers, Concert, Rain, Snow, Vitaliy, Artem, Slava, Juneau, Woods,
- Apply your method to each of our test cases
- Upload the alpha and foreground sequences to any file-sharing service. We kindly ask you to maintain these naming and directory-structure conventions. If your method doesn't explicitly produce the foreground images you can skip uploading them; in this case, we will generate them using method proposed in [7].
- Fill in this form to provide information about your method
Contact us by email with any questions or suggestions at questions@videomatting.com.
Cite Us
To refer to our evaluation or test sequences in your work cite our paper [3].
@inproceedings{Erofeev2015, title={Perceptually Motivated Benchmark for Video Matting}, author={Mikhail Erofeev and Yury Gitman and Dmitriy Vatolin and Alexey Fedorov and Jue Wang}, year={2015}, month={September}, pages={99.1-99.12}, articleno={99}, numpages={12}, booktitle={Proceedings of the British Machine Vision Conference (BMVC)}, publisher={BMVA Press}, doi={10.5244/C.29.99}, isbn={1-901725-53-7}, url={https://dx.doi.org/10.5244/C.29.99} }
Evaluation
Rating
Trimap available for each frame | ||||||||||||
year | rank | city | rain | concert | flowers | snow | Slava | Vitaliy | Artem | juneau | woods | |
---|---|---|---|---|---|---|---|---|---|---|---|---|
Bayesian Matting [2] | 2001 | 12.6 | 58.7614 | 51.6612 | 10.798 | 144.5313 | 30.3613 | 53.2813 | 43.7013 | 70.5014 | 96.6912 | 140.1214 |
Robust Matting [12] | 2007 | 8.2 | 37.5410 | 26.154 | 9.614 | 88.157 | 18.528 | 33.3110 | 34.157 | 46.9111 | 78.5611 | 80.4110 |
Refine Edge [15] | 2013 | 9.9 | 37.7811 | 38.4810 | 8.642 | 129.5611 | 27.4712 | 31.369 | 36.6511 | 38.758 | 102.0213 | 94.5012 |
Closed Form [7] | 2008 | 7.9 | 31.346 | 31.807 | 10.9610 | 149.4814 | 15.865 | 23.677 | 30.155 | 30.605 | 69.949 | 82.1311 |
Learning Based [13] | 2009 | 5.9 | 29.984 | 33.088 | 10.849 | 134.8212 | 15.234 | 23.355 | 31.466 | 30.203 | 58.384 | 56.604 |
Nonlocal mattings [5] | 2011 | 14.7 | 54.2113 | 92.3715 | 33.8615 | 243.4716 | 45.1315 | 64.0515 | 61.7515 | 67.3013 | 135.0815 | 171.7915 |
Shared Matting [4] | 2010 | 8.9 | 36.799 | 46.7211 | 10.307 | 87.646 | 24.4110 | 35.1811 | 35.269 | 45.2310 | 68.528 | 67.178 |
Comprehensive Samplings [10] | 2008 | 7.9 | 34.488 | 30.186 | 12.1011 | 79.604 | 16.176 | 25.028 | 35.258 | 41.059 | 77.7210 | 70.189 |
KNN Matting [1] | 2012 | 13.9 | 67.8315 | 92.2214 | 64.7316 | 112.4610 | 32.9614 | 61.5014 | 50.8714 | 79.2015 | 113.4114 | 123.6313 |
Spectral Mattings [6] | 2012 | 15.5 | 86.3216 | 105.2416 | 13.1512 | 210.0415 | 48.7616 | 65.7216 | 68.8316 | 109.4416 | 157.1816 | 182.8216 |
Sparse Samplings [14] | 2016 | 6.8 | 33.427 | 33.509 | 9.945 | 81.065 | 18.427 | 23.646 | 35.8310 | 38.007 | 65.176 | 64.126 |
Deep Matting [16] | 2016 | 4.6 | 27.333 | 21.403 | 20.3313 | 44.932 | 19.239 | 19.993 | 30.004 | 30.554 | 46.183 | 46.392 |
Self-Adaptive [17] | 2016 | 4.8 | 30.055 | 29.305 | 8.021 | 96.449 | 14.323 | 21.964 | 29.923 | 35.976 | 66.057 | 61.495 |
Information-Flow [20] | 2017 | 10.6 | 50.8112 | 59.6213 | 24.8014 | 94.818 | 25.9311 | 48.8212 | 42.1412 | 55.3212 | 64.025 | 66.567 |
Background Matting [21] | 2020 | 2.5 | 22.772 | 17.532 | 10.066 | 63.613 | 10.802 | 17.152 | 25.191 | 22.272 | 45.442 | 46.853 |
FBA Matting [22] | 2020 | 1.3 | 17.251 | 16.691 | 8.763 | 40.941 | 9.921 | 15.301 | 27.242 | 19.101 | 28.671 | 30.451 |
Bayesian Matting [2] | 2001 | 13.6 | 73.1814 | 67.9213 | 16.0611 | 157.2814 | 39.5914 | 60.3713 | 56.0514 | 85.5315 | 115.8814 | 169.3314 |
Robust Matting [12] | 2007 | 9.5 | 48.1911 | 32.897 | 14.729 | 99.817 | 25.969 | 42.5110 | 38.889 | 59.0811 | 93.7911 | 96.1311 |
Refine Edge [15] | 2013 | 9.6 | 42.2310 | 41.7410 | 10.452 | 127.0211 | 31.4012 | 35.089 | 39.4610 | 42.918 | 108.7112 | 97.7612 |
Closed Form [7] | 2008 | 7.6 | 34.536 | 35.868 | 13.356 | 154.1213 | 21.648 | 27.197 | 33.015 | 35.455 | 81.488 | 92.7510 |
Learning Based [13] | 2009 | 6.2 | 33.445 | 37.639 | 13.235 | 138.3112 | 19.957 | 26.515 | 34.516 | 34.704 | 68.415 | 63.534 |
Nonlocal mattings [5] | 2011 | 14.5 | 55.0413 | 97.8514 | 31.8114 | 242.7516 | 46.9415 | 67.1115 | 65.0915 | 69.0113 | 144.6615 | 176.6115 |
Shared Matting [4] | 2010 | 9.3 | 41.449 | 56.7511 | 14.147 | 96.716 | 27.6611 | 42.9211 | 40.0511 | 58.1710 | 83.029 | 78.918 |
Comprehensive Samplings [10] | 2008 | 6.9 | 38.438 | 28.294 | 14.458 | 83.854 | 16.844 | 27.298 | 37.077 | 47.389 | 89.9610 | 78.457 |
KNN Matting [1] | 2012 | 13.5 | 79.3515 | 98.7115 | 66.2416 | 112.8110 | 34.6013 | 64.1714 | 54.4412 | 84.3814 | 112.1313 | 128.8613 |
Spectral Mattings [6] | 2012 | 15.3 | 99.0516 | 125.2916 | 16.0410 | 214.4215 | 59.6116 | 74.5516 | 77.2516 | 122.8916 | 162.7316 | 188.1616 |
Sparse Samplings [14] | 2016 | 5.2 | 33.204 | 31.925 | 12.344 | 85.525 | 17.665 | 24.644 | 37.788 | 39.876 | 71.056 | 69.945 |
Deep Matting [16] | 2016 | 4.1 | 28.673 | 21.103 | 41.8015 | 44.812 | 16.573 | 21.043 | 30.573 | 30.943 | 46.403 | 50.273 |
Self-Adaptive [17] | 2016 | 6.1 | 34.717 | 32.216 | 11.023 | 101.769 | 18.616 | 26.536 | 31.234 | 40.017 | 75.297 | 70.686 |
Information-Flow [20] | 2017 | 10.5 | 54.6812 | 67.3612 | 27.9313 | 100.448 | 26.9310 | 55.2412 | 55.5413 | 60.1512 | 66.354 | 79.339 |
Background Matting [21] | 2020 | 3.0 | 22.972 | 18.152 | 16.2512 | 68.753 | 10.852 | 17.082 | 24.611 | 22.742 | 43.412 | 46.862 |
FBA Matting [22] | 2020 | 1.1 | 18.901 | 16.871 | 9.851 | 42.481 | 10.161 | 15.641 | 27.352 | 19.851 | 28.801 | 30.951 |
Bayesian Matting [2] | 2001 | 14.1 | 98.1315 | 95.9513 | 38.5414 | 170.7614 | 59.5115 | 80.7215 | 67.6413 | 122.0814 | 135.5914 | 179.2514 |
Robust Matting [12] | 2007 | 10.7 | 67.7113 | 49.029 | 19.877 | 109.799 | 44.1013 | 62.1811 | 50.5410 | 81.4912 | 118.0912 | 115.9611 |
Refine Edge [15] | 2013 | 9.1 | 50.317 | 51.8110 | 15.854 | 125.2511 | 40.7312 | 45.389 | 46.439 | 52.227 | 118.8813 | 103.309 |
Closed Form [7] | 2008 | 7.8 | 45.046 | 41.696 | 24.199 | 159.4713 | 30.099 | 36.858 | 39.204 | 46.975 | 104.078 | 110.2610 |
Learning Based [13] | 2009 | 6.4 | 44.975 | 43.757 | 21.908 | 143.5412 | 26.506 | 33.896 | 40.446 | 45.934 | 90.356 | 81.224 |
Nonlocal mattings [5] | 2011 | 13.7 | 56.7610 | 113.9015 | 33.9713 | 241.3116 | 50.7414 | 77.1414 | 68.3714 | 76.4311 | 159.0815 | 181.1415 |
Shared Matting [4] | 2010 | 9.7 | 55.849 | 79.0911 | 19.786 | 107.458 | 32.9410 | 60.7810 | 51.4711 | 122.3315 | 107.349 | 97.228 |
Comprehensive Samplings [10] | 2008 | 6.9 | 56.7911 | 37.365 | 16.965 | 87.644 | 24.045 | 34.327 | 43.697 | 64.639 | 110.5010 | 93.356 |
KNN Matting [1] | 2012 | 12.7 | 95.8914 | 112.7014 | 66.1216 | 113.1010 | 37.1811 | 71.2613 | 60.0312 | 98.4213 | 110.8711 | 138.6613 |
Spectral Mattings [6] | 2012 | 15.3 | 116.8716 | 157.3716 | 26.1210 | 218.9415 | 76.2216 | 88.7416 | 91.6916 | 148.7016 | 169.3316 | 193.4616 |
Sparse Samplings [14] | 2016 | 4.8 | 43.324 | 34.664 | 14.783 | 89.795 | 19.234 | 29.924 | 44.598 | 51.946 | 85.445 | 81.775 |
Deep Matting [16] | 2016 | 4.0 | 31.293 | 23.162 | 53.0215 | 44.932 | 18.203 | 24.363 | 32.043 | 37.603 | 50.483 | 54.763 |
Self-Adaptive [17] | 2016 | 6.4 | 51.298 | 46.568 | 14.352 | 106.717 | 27.557 | 31.985 | 39.355 | 54.808 | 95.907 | 95.737 |
Information-Flow [20] | 2017 | 10.3 | 66.8612 | 86.0112 | 30.9912 | 104.146 | 29.258 | 64.3212 | 82.9215 | 69.3310 | 69.714 | 117.3812 |
Background Matting [21] | 2020 | 3.0 | 25.772 | 23.623 | 29.3311 | 75.903 | 11.742 | 18.312 | 24.831 | 24.972 | 43.922 | 50.512 |
FBA Matting [22] | 2020 | 1.1 | 22.021 | 19.821 | 11.941 | 44.621 | 11.211 | 17.051 | 28.112 | 22.141 | 29.791 | 33.061 |
Bayesian Matting [2] | 2001 | 13.8 | 35.9314 | 65.2412 | 9.2810 | 104.6815 | 27.6113 | 35.0416 | 46.4113 | 60.3415 | 88.4615 | 121.9115 |
Robust Matting [12] | 2007 | 8.2 | 22.329 | 32.734 | 7.264 | 47.2713 | 16.048 | 18.7210 | 34.477 | 24.968 | 53.4211 | 42.408 |
Refine Edge [15] | 2013 | 7.2 | 14.412 | 44.089 | 6.303 | 40.698 | 18.3610 | 16.536 | 34.938 | 17.834 | 53.4312 | 43.6010 |
Closed Form [7] | 2008 | 6.2 | 16.374 | 41.297 | 7.676 | 35.147 | 14.595 | 16.405 | 32.435 | 17.373 | 51.408 | 47.7112 |
Learning Based [13] | 2009 | 5.8 | 17.296 | 42.998 | 7.888 | 34.035 | 14.444 | 17.078 | 34.376 | 18.375 | 46.024 | 39.464 |
Nonlocal mattings [5] | 2011 | 14.6 | 29.0713 | 110.7614 | 14.1015 | 120.0116 | 29.8215 | 32.6914 | 59.9115 | 36.2214 | 108.4316 | 96.8214 |
Shared Matting [4] | 2010 | 9.2 | 22.7210 | 59.5911 | 7.727 | 45.9212 | 20.3411 | 22.3511 | 38.0211 | 29.5611 | 47.455 | 38.193 |
Comprehensive Samplings [10] | 2008 | 8.9 | 22.7811 | 39.976 | 13.8314 | 33.114 | 15.616 | 17.519 | 37.099 | 25.729 | 53.3310 | 45.0411 |
KNN Matting [1] | 2012 | 13.7 | 36.1615 | 115.0815 | 25.5816 | 44.0910 | 27.8014 | 31.6113 | 56.0814 | 34.8013 | 87.5314 | 75.3613 |
Spectral Mattings [6] | 2012 | 13.9 | 65.5016 | 131.7416 | 4.781 | 103.6114 | 39.0316 | 34.3915 | 63.2116 | 64.4016 | 87.4413 | 122.5316 |
Sparse Samplings [14] | 2016 | 8.9 | 21.788 | 44.8510 | 9.8611 | 34.076 | 16.949 | 16.557 | 37.4010 | 25.8410 | 51.779 | 42.989 |
Deep Matting [16] | 2016 | 4.7 | 16.795 | 25.933 | 13.1213 | 26.742 | 15.647 | 14.583 | 31.853 | 19.256 | 45.033 | 37.272 |
Self-Adaptive [17] | 2016 | 5.8 | 18.767 | 38.895 | 7.375 | 45.7511 | 13.363 | 15.444 | 32.264 | 20.127 | 49.156 | 40.486 |
Information-Flow [20] | 2017 | 10.8 | 28.0012 | 78.5513 | 10.6412 | 41.139 | 22.5612 | 28.7112 | 42.8912 | 29.7512 | 49.897 | 41.447 |
Background Matting [21] | 2020 | 2.4 | 15.003 | 21.442 | 6.252 | 32.743 | 10.532 | 13.402 | 27.791 | 16.292 | 42.392 | 40.105 |
FBA Matting [22] | 2020 | 1.9 | 12.651 | 19.551 | 8.799 | 23.721 | 10.301 | 12.321 | 28.522 | 14.821 | 29.271 | 28.061 |
Bayesian Matting [2] | 2001 | 14.2 | 46.9615 | 77.2712 | 11.3210 | 111.6015 | 33.8515 | 39.5416 | 56.8612 | 75.5016 | 102.7015 | 151.6016 |
Robust Matting [12] | 2007 | 9.3 | 26.3610 | 37.735 | 8.356 | 52.0713 | 19.189 | 21.4510 | 37.088 | 28.5510 | 57.7812 | 47.9510 |
Refine Edge [15] | 2013 | 6.4 | 15.243 | 47.3510 | 6.582 | 40.548 | 19.4010 | 16.775 | 35.746 | 18.244 | 54.1110 | 43.436 |
Closed Form [7] | 2008 | 6.8 | 17.255 | 42.808 | 7.114 | 35.797 | 16.708 | 17.107 | 33.945 | 18.083 | 52.829 | 48.9412 |
Learning Based [13] | 2009 | 5.9 | 18.976 | 45.149 | 7.455 | 34.623 | 16.327 | 18.098 | 36.777 | 19.506 | 48.004 | 39.164 |
Nonlocal mattings [5] | 2011 | 14.1 | 29.6712 | 114.9414 | 14.8413 | 121.9916 | 31.0314 | 33.6614 | 61.3415 | 37.0413 | 117.0216 | 93.1814 |
Shared Matting [4] | 2010 | 9.8 | 24.219 | 68.2511 | 9.008 | 49.9512 | 21.8811 | 24.2511 | 42.7411 | 37.1414 | 51.506 | 39.515 |
Comprehensive Samplings [10] | 2008 | 8.8 | 26.8011 | 35.364 | 15.5914 | 34.805 | 15.734 | 18.299 | 39.1710 | 27.819 | 55.9311 | 48.7311 |
KNN Matting [1] | 2012 | 13.2 | 38.8514 | 120.0815 | 26.5416 | 44.019 | 28.2513 | 31.8313 | 57.5813 | 35.7812 | 85.3714 | 75.3413 |
Spectral Mattings [6] | 2012 | 13.7 | 72.4516 | 155.2116 | 5.031 | 104.9514 | 44.5016 | 35.8415 | 64.9616 | 66.6615 | 85.2313 | 130.3215 |
Sparse Samplings [14] | 2016 | 7.2 | 23.668 | 39.996 | 11.5411 | 35.176 | 16.176 | 16.484 | 38.939 | 27.468 | 51.937 | 44.557 |
Deep Matting [16] | 2016 | 4.3 | 17.214 | 24.743 | 21.0315 | 27.682 | 14.333 | 14.983 | 31.413 | 19.055 | 45.103 | 36.962 |
Self-Adaptive [17] | 2016 | 7.0 | 20.617 | 40.207 | 8.937 | 46.7311 | 15.805 | 17.036 | 32.954 | 21.697 | 52.158 | 44.908 |
Information-Flow [20] | 2017 | 11.1 | 34.1013 | 86.7113 | 13.9512 | 45.9910 | 23.2612 | 31.0012 | 61.2414 | 30.4211 | 49.385 | 44.999 |
Background Matting [21] | 2020 | 2.3 | 14.922 | 22.502 | 6.913 | 34.674 | 10.692 | 13.212 | 27.251 | 16.162 | 40.042 | 38.863 |
FBA Matting [22] | 2020 | 1.9 | 13.671 | 19.981 | 9.629 | 24.321 | 10.471 | 12.411 | 28.712 | 15.111 | 29.151 | 28.291 |
Bayesian Matting [2] | 2001 | 15.2 | 68.3215 | 110.7813 | 29.7016 | 119.5415 | 47.6615 | 52.2416 | 67.6015 | 118.0116 | 98.3415 | 148.1716 |
Robust Matting [12] | 2007 | 9.8 | 32.7211 | 53.128 | 9.855 | 56.8313 | 25.4012 | 27.4110 | 42.738 | 36.4911 | 62.5512 | 52.278 |
Refine Edge [15] | 2013 | 5.8 | 15.823 | 57.5710 | 6.582 | 40.258 | 20.679 | 18.044 | 37.306 | 19.113 | 55.247 | 44.186 |
Closed Form [7] | 2008 | 5.9 | 18.755 | 48.306 | 7.763 | 36.786 | 19.577 | 18.216 | 35.504 | 19.154 | 57.139 | 53.959 |
Learning Based [13] | 2009 | 5.6 | 21.356 | 50.917 | 7.844 | 35.593 | 18.925 | 19.378 | 39.757 | 21.316 | 51.665 | 41.995 |
Nonlocal mattings [5] | 2011 | 13.7 | 31.2510 | 129.1314 | 17.9612 | 123.4916 | 31.7414 | 36.1315 | 62.3213 | 38.8213 | 127.0616 | 91.3714 |
Shared Matting [4] | 2010 | 10.2 | 27.959 | 93.5811 | 10.609 | 55.3212 | 25.0411 | 27.7211 | 48.2811 | 65.0614 | 58.6310 | 41.344 |
Comprehensive Samplings [10] | 2008 | 8.8 | 34.6212 | 42.665 | 16.5311 | 35.694 | 19.236 | 21.269 | 44.3310 | 33.2810 | 60.0111 | 55.6110 |
KNN Matting [1] | 2012 | 12.8 | 42.7114 | 130.1415 | 26.5015 | 43.959 | 29.1913 | 33.0612 | 59.2212 | 38.0412 | 82.4713 | 76.1613 |
Spectral Mattings [6] | 2012 | 13.5 | 81.1416 | 183.7616 | 5.491 | 108.6714 | 54.3116 | 35.8114 | 67.3114 | 70.4515 | 82.7414 | 128.4115 |
Sparse Samplings [14] | 2016 | 6.8 | 27.398 | 41.554 | 12.1510 | 36.957 | 17.044 | 18.195 | 43.949 | 32.678 | 54.816 | 49.267 |
Deep Matting [16] | 2016 | 4.1 | 18.724 | 28.242 | 23.6514 | 28.182 | 15.363 | 16.663 | 31.803 | 20.985 | 47.833 | 36.952 |
Self-Adaptive [17] | 2016 | 7.9 | 25.667 | 56.949 | 10.236 | 49.2910 | 19.978 | 18.717 | 37.005 | 25.317 | 55.898 | 57.1112 |
Information-Flow [20] | 2017 | 11.2 | 41.0913 | 106.6112 | 18.0313 | 49.3011 | 24.5910 | 35.4913 | 94.9216 | 32.869 | 49.224 | 56.9211 |
Background Matting [21] | 2020 | 2.9 | 15.572 | 28.723 | 10.588 | 36.505 | 11.331 | 13.562 | 27.231 | 16.712 | 39.752 | 38.143 |
FBA Matting [22] | 2020 | 1.8 | 14.431 | 23.681 | 10.557 | 24.981 | 11.492 | 12.931 | 29.292 | 15.851 | 29.641 | 29.221 |
Bayesian Matting [2] | 2001 | 13.8 | 1.8514 | 3.5812 | 0.1211 | 13.3314 | 0.7414 | 2.3116 | 1.4713 | 4.5115 | 8.0114 | 16.2315 |
Robust Matting [12] | 2007 | 9.5 | 0.8911 | 0.944 | 0.074 | 4.8213 | 0.259 | 0.8810 | 0.778 | 1.5812 | 3.8912 | 3.9212 |
Refine Edge [15] | 2013 | 8.0 | 0.556 | 1.4810 | 0.062 | 3.9910 | 0.3410 | 0.549 | 0.707 | 0.806 | 3.8011 | 3.169 |
Closed Form [7] | 2008 | 5.7 | 0.493 | 1.358 | 0.098 | 2.884 | 0.195 | 0.356 | 0.613 | 0.614 | 2.756 | 3.3810 |
Learning Based [13] | 2009 | 5.6 | 0.505 | 1.429 | 0.099 | 2.885 | 0.184 | 0.345 | 0.686 | 0.645 | 2.374 | 2.434 |
Nonlocal mattings [5] | 2011 | 14.9 | 1.5613 | 8.1715 | 0.7015 | 20.9716 | 1.1115 | 1.8815 | 2.5116 | 2.3514 | 11.1716 | 13.3714 |
Shared Matting [4] | 2010 | 10.1 | 0.7910 | 2.7311 | 0.097 | 4.5012 | 0.3511 | 1.0012 | 0.9111 | 1.4311 | 2.948 | 2.948 |
Comprehensive Samplings [10] | 2008 | 8.8 | 0.749 | 1.065 | 0.1912 | 3.157 | 0.227 | 0.398 | 0.8610 | 1.249 | 3.6510 | 3.4011 |
KNN Matting [1] | 2012 | 13.2 | 2.2815 | 6.4014 | 1.2216 | 3.478 | 0.6113 | 1.0213 | 1.7314 | 1.9413 | 6.4213 | 7.7313 |
Spectral Mattings [6] | 2012 | 14.5 | 7.0316 | 13.2516 | 0.086 | 15.3615 | 2.0916 | 1.8414 | 2.4515 | 6.5916 | 8.7115 | 19.9716 |
Sparse Samplings [14] | 2016 | 7.9 | 0.668 | 1.307 | 0.1110 | 3.096 | 0.258 | 0.387 | 0.869 | 1.048 | 2.949 | 2.767 |
Deep Matting [16] | 2016 | 4.6 | 0.504 | 0.483 | 0.3214 | 1.152 | 0.226 | 0.253 | 0.635 | 0.573 | 1.843 | 1.843 |
Self-Adaptive [17] | 2016 | 5.6 | 0.587 | 1.156 | 0.061 | 4.4311 | 0.163 | 0.334 | 0.624 | 0.877 | 2.917 | 2.666 |
Information-Flow [20] | 2017 | 10.2 | 1.1012 | 3.9413 | 0.2713 | 3.909 | 0.4112 | 0.9711 | 1.1012 | 1.3810 | 2.465 | 2.615 |
Background Matting [21] | 2020 | 2.3 | 0.382 | 0.322 | 0.085 | 2.113 | 0.102 | 0.202 | 0.431 | 0.352 | 1.622 | 1.822 |
FBA Matting [22] | 2020 | 1.3 | 0.231 | 0.301 | 0.073 | 0.921 | 0.091 | 0.151 | 0.502 | 0.261 | 0.771 | 0.871 |
Bayesian Matting [2] | 2001 | 14.6 | 3.1215 | 5.0713 | 0.2211 | 16.6515 | 1.3415 | 3.0316 | 2.5013 | 11.6016 | 18.0316 | 25.6316 |
Robust Matting [12] | 2007 | 10.7 | 1.3411 | 1.266 | 0.157 | 6.1913 | 0.5012 | 1.4212 | 0.9910 | 2.3812 | 5.2112 | 5.4412 |
Refine Edge [15] | 2013 | 7.0 | 0.665 | 1.6210 | 0.081 | 3.959 | 0.429 | 0.639 | 0.766 | 0.926 | 4.0510 | 3.165 |
Closed Form [7] | 2008 | 6.1 | 0.614 | 1.468 | 0.114 | 3.245 | 0.278 | 0.457 | 0.695 | 0.804 | 3.296 | 3.9310 |
Learning Based [13] | 2009 | 5.5 | 0.666 | 1.599 | 0.115 | 3.224 | 0.266 | 0.424 | 0.787 | 0.845 | 2.965 | 2.874 |
Nonlocal mattings [5] | 2011 | 14.3 | 1.6213 | 8.9315 | 0.7814 | 21.6216 | 1.2214 | 2.0614 | 2.7715 | 2.5013 | 13.5215 | 12.3014 |
Shared Matting [4] | 2010 | 10.8 | 1.039 | 3.5111 | 0.179 | 5.7712 | 0.4811 | 1.5113 | 1.6211 | 2.5714 | 4.049 | 3.789 |
Comprehensive Samplings [10] | 2008 | 8.9 | 1.0910 | 0.954 | 0.2612 | 3.638 | 0.277 | 0.488 | 0.988 | 1.7110 | 4.6511 | 4.4211 |
KNN Matting [1] | 2012 | 12.3 | 3.0514 | 6.8314 | 1.3216 | 3.496 | 0.7013 | 1.1111 | 1.9612 | 2.2811 | 6.3013 | 8.5513 |
Spectral Mattings [6] | 2012 | 14.0 | 9.3616 | 20.5416 | 0.113 | 16.5514 | 3.1516 | 2.3615 | 2.9116 | 7.8915 | 8.8714 | 22.4415 |
Sparse Samplings [14] | 2016 | 7.2 | 0.798 | 1.185 | 0.1810 | 3.527 | 0.265 | 0.446 | 0.989 | 1.298 | 3.397 | 3.367 |
Deep Matting [16] | 2016 | 4.1 | 0.563 | 0.473 | 1.0515 | 1.182 | 0.193 | 0.273 | 0.673 | 0.573 | 1.903 | 1.993 |
Self-Adaptive [17] | 2016 | 6.7 | 0.767 | 1.347 | 0.126 | 4.9411 | 0.244 | 0.435 | 0.694 | 1.127 | 3.668 | 3.628 |
Information-Flow [20] | 2017 | 10.0 | 1.6212 | 4.6112 | 0.4213 | 4.7210 | 0.4610 | 1.1010 | 2.6014 | 1.549 | 2.564 | 3.196 |
Background Matting [21] | 2020 | 2.6 | 0.392 | 0.372 | 0.168 | 2.523 | 0.112 | 0.202 | 0.421 | 0.352 | 1.512 | 1.852 |
FBA Matting [22] | 2020 | 1.2 | 0.281 | 0.311 | 0.092 | 0.971 | 0.091 | 0.161 | 0.512 | 0.281 | 0.781 | 0.921 |
Bayesian Matting [2] | 2001 | 15.4 | 6.7115 | 11.0914 | 3.9716 | 19.8815 | 3.0015 | 5.1816 | 4.0815 | 37.4516 | 16.5016 | 25.0816 |
Robust Matting [12] | 2007 | 11.7 | 2.4012 | 2.5410 | 0.237 | 7.6513 | 1.1813 | 2.7014 | 1.5710 | 4.1613 | 7.2613 | 7.1512 |
Refine Edge [15] | 2013 | 5.7 | 0.814 | 2.188 | 0.121 | 3.917 | 0.548 | 0.849 | 0.905 | 1.144 | 4.517 | 3.394 |
Closed Form [7] | 2008 | 6.1 | 0.885 | 1.706 | 0.248 | 3.646 | 0.466 | 0.727 | 0.854 | 1.205 | 4.366 | 5.218 |
Learning Based [13] | 2009 | 5.5 | 1.036 | 1.887 | 0.215 | 3.625 | 0.435 | 0.614 | 1.007 | 1.326 | 4.275 | 3.745 |
Nonlocal mattings [5] | 2011 | 13.2 | 1.7610 | 11.8515 | 1.1013 | 22.2516 | 1.2914 | 2.4612 | 3.0013 | 2.8810 | 16.3615 | 12.2014 |
Shared Matting [4] | 2010 | 10.8 | 1.749 | 6.0211 | 0.279 | 7.4612 | 0.8511 | 2.7013 | 2.2611 | 7.6514 | 6.3511 | 5.057 |
Comprehensive Samplings [10] | 2008 | 9.1 | 2.2811 | 1.555 | 0.3210 | 3.978 | 0.547 | 0.738 | 1.358 | 3.0111 | 6.5912 | 6.1111 |
KNN Matting [1] | 2012 | 11.4 | 4.2414 | 7.8413 | 1.3214 | 3.524 | 0.8812 | 1.3810 | 2.3412 | 3.0412 | 6.1710 | 9.8713 |
Spectral Mattings [6] | 2012 | 13.8 | 13.1716 | 31.3816 | 0.183 | 18.1614 | 5.3316 | 2.7815 | 3.7414 | 10.2415 | 9.1914 | 22.9215 |
Sparse Samplings [14] | 2016 | 6.7 | 1.377 | 1.324 | 0.236 | 4.029 | 0.354 | 0.655 | 1.439 | 2.199 | 4.678 | 4.666 |
Deep Matting [16] | 2016 | 4.0 | 0.643 | 0.602 | 1.3415 | 1.222 | 0.223 | 0.373 | 0.753 | 0.783 | 2.273 | 2.213 |
Self-Adaptive [17] | 2016 | 8.0 | 1.428 | 2.499 | 0.194 | 5.4811 | 0.559 | 0.666 | 0.996 | 1.948 | 5.209 | 6.0410 |
Information-Flow [20] | 2017 | 10.4 | 2.6713 | 6.2512 | 0.7112 | 5.3210 | 0.5710 | 1.5411 | 5.3216 | 1.927 | 2.774 | 5.839 |
Background Matting [21] | 2020 | 3.0 | 0.472 | 0.653 | 0.4911 | 2.983 | 0.122 | 0.222 | 0.431 | 0.412 | 1.562 | 2.022 |
FBA Matting [22] | 2020 | 1.2 | 0.351 | 0.431 | 0.142 | 1.051 | 0.121 | 0.191 | 0.532 | 0.341 | 0.821 | 1.061 |
- city
- rain
- concert
- flowers
- snow
- Slava
- Vitaliy
- Artem
- juneau
- woods
Integral Plots
Subjective comparison
We carried out subjective comparison of 13 matting methods using Subjectify.us platform. We applied matting methods to videos from our dataset and then uploaded videos containing extracted foreground objects and ground-truth sequences to Subjectify.us. The platform hired study participants and showed them these videos in pairwise fashion. For each pair, participants were asked to choose the video with better visual quality or indicate that they are approximately equal. Each study participant compared 30 pairs including 4 hidden quality-control comparisons between ground truth and a low-quality method; answers of 23 participants were rejected, since they failed at least one quality-control question. In total 10556 answers from 406 participants were collected. Bradley-Terry [18] and Crowd Bradley-Terry [19] models were used to convert pairwise comparisons to subjective ranks. The study report generated by the platform is shown below.
Multidimensional Analysis
References
[1] | Qifeng Chen, Dingzeyu Li, and Chi-Keung Tang. KNN matting. Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 35(9):2175–2188, 2013. [ doi , project page ] |
[2] | Yung-Yu Chuang, Brian Curless, David H. Salesin, and Richard Szeliski. A bayesian approach to digital matting. In Computer Vision and Pattern Recognition (CVPR), volume 2, pages II-264–II-271, 2001. [ doi , project page , code ] |
[3] | Mikhail Erofeev, Yury Gitman, Dmitriy Vatolin, Alexey Fedorov, Jue Wang. Perceptually Motivated Benchmark for Video Matting. British Machine Vision Conference (BMVC), pages 99.1–99.12, 2015. [ doi , pdf , project page ] |
[4] | Eduardo S.L. Gastal and Manuel M. Oliveira. Shared sampling for real-time alpha matting. Computer Graphics Forum, 29(2):575–584, 2010. [ project page ] |
[5] | Philip Lee and Ying Wu. Nonlocal matting. In Computer Vision and Pattern Recognition (CVPR), pages 2193–2200, 2011. [ code ] |
[6] | A. Levin, A. Rav Acha, and D. Lischinski. Spectral matting. Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 30(10):1699–1712, 2008. [ doi , project page ] |
[7] | Anat Levin, Dani Lischinski, and Yair Weiss. A closed-form solution to natural image matting. Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 30(2):228–242, 2008. [ doi , code ] |
[8] | Christoph Rhemann, Carsten Rother, Jue Wang, Margrit Gelautz, Pushmeet Kohli, and Pamela Rott. A perceptually motivated online benchmark for image matting. In Computer Vision and Pattern Recognition (CVPR), pages 1826–1833, 2009. [ doi ] |
[9] | E. Shahrian and D. Rajan. Weighted color and texture sample selection for image matting. In Computer Vision and Pattern Recognition (CVPR), pages 718–725, 2012. [ doi , code ] |
[10] | E. Shahrian, D. Rajan, B. Price, and S. Cohen. Improving image matting using comprehensive sampling sets. In Computer Vision and Pattern Recognition (CVPR), pages 636–643, 2013. [ doi , code ] |
[11] | Karen Symonyan, Sergey Grishin, Dmitriy Vatolin, and Dmitriy Popov. Fast video superresolution via classification. International Conference on Image Processing (ICIP), pages 349–352, 2008. [ doi ] |
[12] | Jue Wang and Michael F. Cohen. Optimized color sampling for robust matting. In Computer Vision and Pattern Recognition (CVPR), pages 1–8, 2007. [ doi , project page ] |
[13] | Yuanjie Zheng and C. Kambhamettu. Learning based digital matting. In International Conference on Computer Vision (ICCV), pages 889–896, 2009. [ doi , code ] |
[14] | Levent Karacan, Aykut Erdem, Erkut Erdem. Alpha Matting with KL-Divergence Based Sparse Sampling, IEEE Transactions on Image Processing, 2017. |
[15] | http://www.adobe.com/en/products/aftereffects.html, Refine Edge tool in Adobe After Effects CC. |
[16] | Ning Xu, Brian Price, Scott Cohen, and Thomas Huang. Deep Matting. In Computer Vision and Pattern Recognition (CVPR), 2017 |
[17] | Guangying Cao, Jianwei Li, Xiaowu Chen, Zhiqiang He. Patch-based self-adaptive matting for high-resolution image and video, The Visual Computer, 1-15, 2017. |
[18] | Ralph Allan Bradley and Milton E Terry. Rank analysis of incomplete block designs: I. the method of paired comparisons. Biometrika, 39(3/4):324–345, 1952. |
[19] | Chen, Xi, et al. Pairwise ranking aggregation in a crowdsourced setting. Proceedings of the sixth ACM international conference on Web search and data mining. ACM, 2013. |
[20] | Yagiz Aksoy, Tunc Ozan Aydin and Marc Pollefeys. Designing Effective Inter-Pixel Information Flow for Natural Image Matting. In Computer Vision and Pattern Recognition (CVPR), 2017. [ doi , code ] |
[21] | Matting with Background Estimation: A Novel Method for Extracting Clean Foreground, IEEE Transaction on Image Processing 2020 (anonymous submission) |
[22] | F, B, Alpha Matting. Anonymous ECCV 2020 submission #6826 |