#### DMCA

## SparseFlow: Sparse Matching for Small to Large Displacement Optical Flow

### Citations

8932 | Distinctive image features from scale-invariant keypoints
- Lowe
- 2004
(Show Context)
Citation Context ...splacements and for wrongfully proposed matches. In the context of scene correspondence, the SIFT-flow [19] and PatchMatch [5] algorithms use descriptors or small patches. Xu et al.[37] combines SIFT =-=[21]-=- and PatchMatch [5] matching for refined flow level initialization with excellent performance at the expense of computation costs. Leordeanu et al. [18] extend coarse matching to dense matching by enf... |

3729 | Histogram of Oriented Gradients for Human Detection
- Dalal, Triggs
- 2005
(Show Context)
Citation Context ...ome arbitrarily large displacements, a strength thus incorporated into the variational optical flow methods. Most current matching approaches are based on descriptors with a square support (e.g. HOGs =-=[12]-=-), that are invariant only to similarities. Yet, exactly under the conditions where large displacements need to be bridged, this level of invariance may be insufficient [9]. Weinzaepfel et al. [33] im... |

3592 | Compressed sensing
- Donoho
- 2006
(Show Context)
Citation Context ... matching’ solution is inspired by deep convolutional nets [17], has 6 layers, and interleaves convolutions and max-pooling. We propose a novel matching process that is inspired by compressed sensing =-=[13]-=-. Thus, we work under a sparsity assumption. The pixels are described by their surrounding blocks of pixel intensities. A pixel can then be sparsely decomposed over a pool of pixels from a target imag... |

2408 | Nonlinear Dimensionality Reduction by Locally Linear Embedding”, Science 290
- Roweis, Saul
(Show Context)
Citation Context ... shown in [30], with a tolerance θ = 0.05 we recover the coefficients up to 0.95 and need K NN iterations of the INN algorithm, where K = d− log(1− θ) log(1 + λ) e. (6) Locally Linear Embedding (LLE) =-=[26]-=- encodes a sample y over its neighborhood NX(y) of size K in X, zi ∈ NX(y). The sparsity is directly imposed by restricting the decomposition to the local neighborhood of samples. ĉ = argmin c ‖y− K∑... |

2402 | Determining optical flow
- Horn, Schunck
- 1981
(Show Context)
Citation Context ...paper in Section 6. 2. Related work Large displacement in optical flow estimation. The state-of-the-art in optical flow is represented by the variational methods. The seminal work of Horn and Schunck =-=[15]-=- has been improved repeatedly over the years [25, 6, 10, 24, 34, 28, 4, 32]. Brox et al. [8] combine many of these improvements into a variational approach. The problem is formulated as an energy mini... |

1523 | Gradient-based learning applied to document recognition
- LeCun, Bottou, et al.
- 1998
(Show Context)
Citation Context ...prove the descriptor matching by not only increasing the density of matched points, but also by catering for deformable matching. Their ‘deep matching’ solution is inspired by deep convolutional nets =-=[17]-=-, has 6 layers, and interleaves convolutions and max-pooling. We propose a novel matching process that is inspired by compressed sensing [13]. Thus, we work under a sparsity assumption. The pixels are... |

933 | Robust face recognition via sparse representation
- Wright, Yang, et al.
- 2009
(Show Context)
Citation Context ...can then be sparsely decomposed over a pool of pixels from a target image. This sparse decomposition formulation is able to cope with high image corruptions and deformations as shown by Wright et al. =-=[36]-=- for face recognition. The dominant pixel in the decomposition is likely to be the correspondence in the target image. We call this process of sparse coding and correspondence selection sparse matchin... |

705 | Lucas-Kanade 20 Years On: A unifying framework
- BAKER, MATTHEWS
- 2004
(Show Context)
Citation Context ...Comparison of matching algorithms We compare our sparse matching directly with the deep matching code provided by its authors [33], and indirectly with diverse state-of-the-art algorithms: KLT tracks =-=[3]-=-, sparse SIFT matching [21] (here SIFT-NN), dense HOG matching with uniqueness as in LDOF [9] (here HOG-NN). For the quantitative results as we report them, we adhere to the setup proposed by Weinzaep... |

669 |
The robust estimation of multiple motions: parametric and piecewise smooth flow fields
- Black, Anandan
- 1996
(Show Context)
Citation Context ...ities (outliers, occlusions, motion discontinuities), appearance changes (illumination, chromacity, deformations), and large displacements. While we have efficient approaches for the first two issues =-=[6, 24]-=-, how to handle large displacements to a large extent still is an open problem, despite the recent endeavors [35, 27, 9, 37, 33, 7, 18]. The seminal work of Brox and Malik [9] shows that a variational... |

509 | High accuracy optical flow estimation based on a theory for warping
- Brox, Bruhn, et al.
- 2004
(Show Context)
Citation Context ...of-the-art in optical flow is represented by the variational methods. The seminal work of Horn and Schunck [15] has been improved repeatedly over the years [25, 6, 10, 24, 34, 28, 4, 32]. Brox et al. =-=[8]-=- combine many of these improvements into a variational approach. The problem is formulated as an energy minimization represented by Euler-Lagrange equations, finally reduced to solving a sequence of l... |

407 | A database and evaluation methodology for optical flow
- Baker, Scharstein, et al.
- 2011
(Show Context)
Citation Context ...ement in optical flow estimation. The state-of-the-art in optical flow is represented by the variational methods. The seminal work of Horn and Schunck [15] has been improved repeatedly over the years =-=[25, 6, 10, 24, 34, 28, 4, 32]-=-. Brox et al. [8] combine many of these improvements into a variational approach. The problem is formulated as an energy minimization represented by Euler-Lagrange equations, finally reduced to solvin... |

337 |
Gool, “A comparison of affine region detectors,”
- Mikolajczyk, Tuytelaars, et al.
- 2005
(Show Context)
Citation Context ... descriptors and matching are the two steps usually employed in matching images. While, initially, the descriptors of choice were extracted sparsely, invariant under scaling or affine transformations =-=[23]-=-, the recent trend in optical flow estimation, is to densely extract rigid (square) descriptors from local frames [31, 9, 19]. The descriptor matching is usually reduced to a (reciprocal) nearest neig... |

195 | M.J.: Secrets of optical flow estimation and their principles
- Sun, Roth, et al.
(Show Context)
Citation Context ...ement in optical flow estimation. The state-of-the-art in optical flow is represented by the variational methods. The seminal work of Horn and Schunck [15] has been improved repeatedly over the years =-=[25, 6, 10, 24, 34, 28, 4, 32]-=-. Brox et al. [8] combine many of these improvements into a variational approach. The problem is formulated as an energy minimization represented by Euler-Lagrange equations, finally reduced to solvin... |

169 |
Large displacement optical flow: Descriptor matching in variational motion estimation.
- Brox, Malik
- 2011
(Show Context)
Citation Context ...nd large displacements. While we have efficient approaches for the first two issues [6, 24], how to handle large displacements to a large extent still is an open problem, despite the recent endeavors =-=[35, 27, 9, 37, 33, 7, 18]-=-. The seminal work of Brox and Malik [9] shows that a variational approach can better handle large displacements when a descriptor matching term is added. The idea is to guide the variational optical ... |

124 | SIFT Flow: Dense Correspondence across Scenes and its Applications” [2] Dizan Vasquez & Thierry Fraichard Inria Rhône-Alpes & Lab. Gravir, Grenoble (FR) - “Motion Prediction for Moving Objects: A Statistical Approach”.
- Liu, IEEE, et al.
- 2011
(Show Context)
Citation Context ...onent to the variational formulation can harm the performance, especially in places with small displacements and for wrongfully proposed matches. In the context of scene correspondence, the SIFT-flow =-=[19]-=- and PatchMatch [5] algorithms use descriptors or small patches. Xu et al.[37] combines SIFT [21] and PatchMatch [5] matching for refined flow level initialization with excellent performance at the ex... |

108 | Highly accurate optic flow computation with theoretically justified warping. IJCV
- Papenberg, Bruhn, et al.
- 2005
(Show Context)
Citation Context ...ities (outliers, occlusions, motion discontinuities), appearance changes (illumination, chromacity, deformations), and large displacements. While we have efficient approaches for the first two issues =-=[6, 24]-=-, how to handle large displacements to a large extent still is an open problem, despite the recent endeavors [35, 27, 9, 37, 33, 7, 18]. The seminal work of Brox and Malik [9] shows that a variational... |

97 | A Fast Local Descriptor for Dense Matching. In
- Tola, Lepetit, et al.
- 2008
(Show Context)
Citation Context ...choice were extracted sparsely, invariant under scaling or affine transformations [23], the recent trend in optical flow estimation, is to densely extract rigid (square) descriptors from local frames =-=[31, 9, 19]-=-. The descriptor matching is usually reduced to a (reciprocal) nearest neighbor operation [21, 5, 9]. Important exceptions are the recent works of Leordeanu et al. [18] (enforcing affine constraints) ... |

77 |
The generalized patchmatch correspondence algorithm.
- Barnes, Shechtman, et al.
- 2010
(Show Context)
Citation Context ...onal formulation can harm the performance, especially in places with small displacements and for wrongfully proposed matches. In the context of scene correspondence, the SIFT-flow [19] and PatchMatch =-=[5]-=- algorithms use descriptors or small patches. Xu et al.[37] combines SIFT [21] and PatchMatch [5] matching for refined flow level initialization with excellent performance at the expense of computatio... |

72 | Motion detail preserving optical flow estimation
- Xu, Jia, et al.
- 2010
(Show Context)
Citation Context ...nd large displacements. While we have efficient approaches for the first two issues [6, 24], how to handle large displacements to a large extent still is an open problem, despite the recent endeavors =-=[35, 27, 9, 37, 33, 7, 18]-=-. The seminal work of Brox and Malik [9] shows that a variational approach can better handle large displacements when a descriptor matching term is added. The idea is to guide the variational optical ... |

70 |
Determination of optical flow and its discontinuities using non-linear diffusion.
- Proesmans, Gool, et al.
- 1994
(Show Context)
Citation Context ...ement in optical flow estimation. The state-of-the-art in optical flow is represented by the variational methods. The seminal work of Horn and Schunck [15] has been improved repeatedly over the years =-=[25, 6, 10, 24, 34, 28, 4, 32]-=-. Brox et al. [8] combine many of these improvements into a variational approach. The problem is formulated as an energy minimization represented by Euler-Lagrange equations, finally reduced to solvin... |

61 | Anisotropic Huber-L1 optical flow. - Werlberger, Trobin, et al. - 2009 |

44 | Towards ultimate motion estimation: combining highest accuracy with real-time performance - Bruhn, Weickert - 2005 |

44 | A naturalistic open source movie for optical flow evaluation.
- Butler, Wulff, et al.
- 2012
(Show Context)
Citation Context ... variational optical flow methods (SparseFlow, SparseFlowFused) inherit the precision and robustness to large displacements offered by sparse matching, providing top performance on MPI-Sintel dataset =-=[11]-=- and KITTI dataset [14]. The remainder is organized as follows. First, we review recent related work in Section 2. Then we introduce the sparse matching algorithm in Section 3. Section 4 describes our... |

40 |
Large displacement optical flow computation without warping
- Steinbrücker, Pock
- 2009
(Show Context)
Citation Context ...nd large displacements. While we have efficient approaches for the first two issues [6, 24], how to handle large displacements to a large extent still is an open problem, despite the recent endeavors =-=[35, 27, 9, 37, 33, 7, 18]-=-. The seminal work of Brox and Malik [9] shows that a variational approach can better handle large displacements when a descriptor matching term is added. The idea is to guide the variational optical ... |

33 |
DeepFlow: Large displacement optical flow with deep matching
- Weinzaepfel, Revaud, et al.
(Show Context)
Citation Context ...hat are the most successful at handling large displacements blend sparse correspondences from a matching algorithm with an optimization that refines the optical flow. We follow the scheme of DeepFlow =-=[33]-=-. We first extract sparse pixel correspondences by means of a matching procedure and then apply a variational approach to obtain a refined optical flow. In our approach, coined ‘SparseFlow’, the novel... |

21 | A feature-based approach for dense segmentation and estimation of large disparity motion. IJCV
- Wills, Agarwal, et al.
- 2006
(Show Context)
Citation Context |

9 | Gool. Iterative nearest neighbors for classification and dimensionality reduction
- Timofte, Van
(Show Context)
Citation Context ... ŵ = argmin w ‖y −Xw‖22 + λ‖w‖1, (3) where y is the query (pixel feature in our case), X the pool, w are the coefficients, λ is the regulatory parameter. The Iterative Nearest Neighbors (INN) method =-=[29, 30]-=- combines the power of SR with the computational simplicity of NN by means of a constrained decomposition: {ŝ}Ki=1 = argmin {s}Ki=1 ‖y − K∑ i=1 λ (1 + λ)i si‖2 (4) where λ is the regulatory parameter... |

9 | An evaluation of data costs for optical flow
- Vogel, Roth, et al.
- 2013
(Show Context)
Citation Context |

8 |
Locally affine sparseto-dense matching for motion and occlusion estimation
- Leordeanu, Zanfir, et al.
- 2013
(Show Context)
Citation Context |

7 | A general dense image matching framework combining direct and feature-based costs
- Braux-Zin, Dupont, et al.
(Show Context)
Citation Context |

4 | Gool. Iterative nearest neighbors
- Timofte, Van
(Show Context)
Citation Context ... ŵ = argmin w ‖y −Xw‖22 + λ‖w‖1, (3) where y is the query (pixel feature in our case), X the pool, w are the coefficients, λ is the regulatory parameter. The Iterative Nearest Neighbors (INN) method =-=[29, 30]-=- combines the power of SR with the computational simplicity of NN by means of a constrained decomposition: {ŝ}Ki=1 = argmin {s}Ki=1 ‖y − K∑ i=1 λ (1 + λ)i si‖2 (4) where λ is the regulatory parameter... |

3 |
Vision meets robotics: The kitti dataset. IJRR
- Geiger, Lenz, et al.
- 2013
(Show Context)
Citation Context ...ow methods (SparseFlow, SparseFlowFused) inherit the precision and robustness to large displacements offered by sparse matching, providing top performance on MPI-Sintel dataset [11] and KITTI dataset =-=[14]-=-. The remainder is organized as follows. First, we review recent related work in Section 2. Then we introduce the sparse matching algorithm in Section 3. Section 4 describes our variational optical fl... |

1 |
SPAMS: SPArse Modeling Software, v2.4. http://spamsdevel.gforge.inria.fr/downloads.html
- Mairal, Bach, et al.
- 2014
(Show Context)
Citation Context ... minimum corner score to 1. In this way, we extract only a few thousands pixel descriptors per image. For obtaining the linear decomposition we considered SR (with the lasso solver from SPAMS library =-=[22]-=-), INN (Matlab solver provided by the authors [30]), and LLE (with Matlab codes based on [26]). Our choice of parameters is λ = 0.1 for SR, λ = 0.25 for INN, and 7 the number of nearest neighbors for ... |