Graph Cut Based Inference with Co-occurrence Statistics

Ladický, Ľubor; Russell, Chris; Kohli, Pushmeet; Torr, Philip H. S.

doi:10.1007/978-3-642-15555-0_18

Cited by 247 publications

(378 citation statements)

References 25 publications

Supporting

Mentioning

372

Contrasting

Order By: Relevance

“…A recent formulation [8] which has been proposed attempts to capture such information in a global co-occurrence potential defined over the entire image clique c I (generalization to arbitrary cliques is also possible) as:…”

Section: Inference In Models With Higher-order Termsmentioning

confidence: 99%

“…For object class segmentation, the importance of enforcing label consistency over homogeneous regions has been demonstrated using P n -Potts models [7], and co-occurrence relations between classes at the image level have also been shown to provide important priors for segmentation [8]. For stereo and optical flow, second-order priors have proved to be effective [9], as have higher-order image priors for denoising [10].…”

Section: Introductionmentioning

confidence: 99%

“…Specifically, we show how to encode (a) a broad class of local pattern-based potentials (as introduced in [11]), which include P nPotts models and second-order smoothness priors, and (b) global potentials representing co-occurrence relationships between labels as in [8,12]. We assume a base-layer MRF with full connectivity and weighted Gaussian edge potentials as in [4].…”

Section: Introductionmentioning

confidence: 99%

“…Our approach allows us to apply bilateral filter-based inference to a wide range of models with complex higher-order structure. We demonstrate the approach on two such models, first a model for joint stereo and object class labeling as in [13], and second a model for object class segmentation with cooccurrence priors as in [8]. In the case of joint stereo and object labeling, in addition to demonstrating fast inference with higher-order terms, we show how cost-volume filtering can be applied in the product label-space to generate informative disparity potentials, and more generally how our method provides an efficient approach to inference in such product label-spaces.…”

Section: Introductionmentioning

confidence: 99%

See 3 more Smart Citations

Filter-Based Mean-Field Inference for Random Fields with Higher-Order Terms and Product Label-Spaces

Warrell²,

2014

Self Cite

View full text Add to dashboard Cite

Abstract. Recently, a number of cross bilateral filtering methods have been proposed for solving multi-label problems in computer vision, such as stereo, optical flow and object class segmentation that show an order of magnitude improvement in speed over previous methods. These methods have achieved good results despite using models with only unary and/or pairwise terms. However, previous work has shown the value of using models with higher-order terms e.g. to represent label consistency over large regions, or global co-occurrence relations. We show how these higher-order terms can be formulated such that filter-based inference remains possible. We demonstrate our techniques on joint stereo and object labeling problems, as well as object class segmentation, showing in addition for joint object-stereo labeling how our method provides an efficient approach to inference in product label-spaces. We show that we are able to speed up inference in these models around 10-30 times with respect to competing graph-cut/move-making methods, as well as maintaining or improving accuracy in all cases. We show results on PascalVOC-10 for object class segmentation, and Leuven for joint object-stereo labeling.

show abstract

Section: Inference In Models With Higher-order Termsmentioning

confidence: 99%

Section: Introductionmentioning

confidence: 99%

Section: Introductionmentioning

confidence: 99%

Section: Introductionmentioning

confidence: 99%

See 2 more Smart Citations

Filter-Based Mean-Field Inference for Random Fields with Higher-Order Terms and Product Label-Spaces

Warrell²,

2014

Self Cite

View full text Add to dashboard Cite

show abstract

“…This technique was inspired by the concave-convex procedure for continuous functions [9]. Our algorithm changes the underlying graph in a way different from earlier submodular-supermodular techniques [10,11]. For example, besides changing the edge capacities, we also replace certain edges modifying the graph's connectivity.…”

Section: Introductionmentioning

confidence: 99%

Hausdorff Distance Constraint for Multi-surface Segmentation

Schmidt

Boykov

2012

Computer Vision – ECCV 2012

View full text Add to dashboard Cite

Abstract. It is well known that multi-surface segmentation can be cast as a multi-labeling problem. Different segments may belong to the same semantic object which may impose various inter-segment constraints [1]. In medical applications, there are a lot of scenarios where upper bounds on the Hausdorff distances between subsequent surfaces are known. We show that incorporating these priors into multi-surface segmentation is potentially NP-hard. To cope with this problem we develop a submodularsupermodular procedure that converges to a locally optimal solution well-approximating the problem. While we cannot guarantee global optimality, only feasible solutions are considered during the optimization process. Empirically, we get useful solutions for many challenging medical applications including MRI and ultrasound images.

show abstract

Optimization and Machine Learning

2022

View full text Add to dashboard Cite

Optimization and machine learning are both extremely active research topics. In this thesis, we explore problems at the intersection of the two fields. In particular, we will develop two main ideas.First, optimization can be used to improve machine learning. We illustrate this idea by considering computer vision tasks that are modelled with dense conditional random fields. Existing solvers for these models are either slow or inaccurate. We show that, by introducing a specialized solver based on proximal minimization and fast filtering, these models can be solved both quickly and accurately. Similarly, we introduce a specialized linear programming solver for block bounded problems, a common class of problems encountered in machine learning. This solver is efficient, easy to tune and simple to integrate inside larger machine learning algorithms.Second, machine learning can be used to improve optimization, in particular for NP-hard problems. For problems solved by using hand-tuned heuristics, machine learning can be used to discover and improve these heuristics. We show that, for the problem of super-optimization, a better heuristic to explore the space of programs can be learnt using reinforcement learning. For problems where no such heuristics exist, machine learning can be used to get an approximate solution of the original problem. We use this idea to tackle the problem of program synthesis by reformulating it as the problem of learning a program that performs the required task. We introduce a new differentiable formulation of the execution and show that the fastest programs can be recovered for simple tasks.

show abstract

Graph Cut Based Inference with Co-occurrence Statistics

Cited by 247 publications

References 25 publications

Filter-Based Mean-Field Inference for Random Fields with Higher-Order Terms and Product Label-Spaces

Filter-Based Mean-Field Inference for Random Fields with Higher-Order Terms and Product Label-Spaces

Hausdorff Distance Constraint for Multi-surface Segmentation

Optimization and Machine Learning

Contact Info

Product

Resources

About