1. Fast and accurate energy minimization for static or time-varying Markov Random Fields (MRFs) Nikos Komodakis (Ecole Centrale Paris) Nikos Paragios (Ecole Centrale Paris) Georgios Tziritas (University of Crete)

2. The MRF optimization problem vertices G = set of objects edges E = object relationships set L = discrete set of labels V p (x p ) = cost of assigning label x p to vertex p (also called single node potential) V pq (x p,x q ) = cost of assigning labels (x p,x q ) to neighboring vertices (p,q) (also called pairwise potential) Find labels that minimize the MRF energy (i.e., the sum of all potentials):

3. MRF optimization in vision MRFs ubiquitous in vision and beyond Have been used in a wide range of problems: segmentationstereo matching optical flowimage restoration image completion object detection & localization... MRF optimization is thus a task of fundamental importance Yet, highly non-trivial, since almost all interesting MRFs are actually NP-hard to optimize Many proposed algorithms (e.g., [Boykov,Veksler,Zabih], [Kolmogorov], [Kohli,Torr], [Wainwright]…)

4. MRF hardness MRF pairwise potential MRF hardness linear exact global optimum arbitrary local optimum metric global optimum approximation Move right in the horizontal axis, But we want to be able to do that efficiently, i.e. fast and remain low in the vertical axis (i.e., still be able to provide approximately optimal solutions)

5. Our contributions to MRF optimization Can handle a very wide class of MRFs General framework for optimizing MRFs based on duality theory of Linear Programming (the Primal-Dual schema) Can guarantee approximately optimal solutions (worst-case theoretical guarantees) Can provide tight certificates of optimality per-instance (per-instance guarantees) Fast-PD Provides significant speed-up for static MRFs Provides significant speed-up for dynamic MRFs

6. Presentation outline The primal-dual schema Applying the schema to MRF optimization Algorithmic properties  Worst-case optimality guarantees  Per-instance optimality guarantees  Computational efficiency for static MRFs  Computational efficiency for dynamic MRFs

7. The primal-dual schema Highly successful technique for exact algorithms. Yielded exact algorithms for cornerstone combinatorial problems: matchingnetwork flow minimum spanning tree minimum branching shortest path... Soon realized that it’s also an extremely powerful tool for deriving approximation algorithms: set coversteiner tree steiner networkfeedback vertex set scheduling...

8. The primal-dual schema  Say we seek an optimal solution x* to the following integer program (this is our primal problem): (NP-hard problem)  To find an approximate solution, we first relax the integrality constraints to get a primal & a dual linear program: primal LP: dual LP:

9. The primal-dual schema Goal: find integral-primal solution x, feasible dual solution y such that their primal-dual costs are “close enough”, e.g., primal cost of solution x primal cost of solution x dual cost of solution y dual cost of solution y cost of optimal integral solution x* cost of optimal integral solution x* Then x is an f * -approximation to optimal solution x*

10. The primal-dual schema sequence of dual costs sequence of primal costs … unknown optimum … The primal-dual schema works iteratively Global effects, through local improvements! Instead of working directly with costs (usually not easy), use RELAXED complementary slackness conditions (easier) Different relaxations of complementary slackness Different approximation algorithms!!!

11. (only one label assigned per vertex) enforce consistency between variables x p,a, x q,b and variable x pq,ab The primal-dual schema for MRFs Binary variables x p,a =1 label a is assigned to node p x pq,ab =1 labels a, b are assigned to nodes p, q x p,a =1 label a is assigned to node p x pq,ab =1 labels a, b are assigned to nodes p, q

12. The primal-dual schema for MRFs During the PD schema for MRFs, it turns out that: each update of primal and dual variables solving max-flow in appropriately constructed graph Max-flow graph defined from current primal-dual pair (x k,y k )  (x k,y k ) defines connectivity of max-flow graph  (x k,y k ) defines capacities of max-flow graph Max-flow graph is thus continuously updated Resulting flows tell us how to update both:  the dual variables, as well as  the primal variables for each iteration of primal-dual schema

13. The primal-dual schema for MRFs Very general framework. Different PD-algorithms by RELAXING complementary slackness conditions differently. Theorem: All derived PD-algorithms shown to satisfy certain relaxed complementary slackness conditions Worst-case optimality properties are thus guaranteed E.g., simply by using a particular relaxation of complementary slackness conditions (and assuming V pq (·,·) is a metric) THEN resulting algorithm shown equivalent to a-expansion! PD-algorithms for non-metric potentials V pq (·,·) as well

14. Per-instance optimality guarantees Primal-dual algorithms can always tell you (for free) how well they performed for a particular instance unknown optimum …… per-instance approx. factor per-instance lower bound (per-instance certificate) per-instance upper bound

15. Computational efficiency (static MRFs) MRF algorithm only in the primal domain (e.g., a-expansion) primal k primal k-1 primal 1 … primal costs dual 1 fixed dual cost gap k STILL BIG Many augmenting paths per max-flow Theorem: primal-dual gap = upper-bound on #augmenting paths (i.e., primal-dual gap indicative of time per max-flow) dual k dual 1 dual k-1 … dual costs gap k primal k primal k-1 primal 1 … primal costs SMALL Few augmenting paths per max-flow MRF algorithm in the primal-dual domain (Fast-PD)

16. Computational efficiency (static MRFs) dramatic decrease always very high Incremental construction of max-flow graphs (recall that max-flow graph changes per iteration) This is possible only because we keep both primal and dual information Our framework provides a principled way of doing this incremental graph construction for general MRFs noisy imagedenoised image

17. Computational efficiency (static MRFs) penguin Tsukuba SRI-tree almost constant dramatic decrease

18. Computational efficiency (dynamic MRFs) Fast-PD can speed up dynamic MRFs [Kohli,Torr] as well (demonstrates the power and generality of our framework) gap primal x dual y SMALL primal x gap dual y SMALL few path augmentations primal x SMALL gap dual 1 fixed dual cost primal x gap LARGE many path augmentations Our framework provides principled (and simple) way to update dual variables when switching between different MRFs Fast-PD algorithm primal-based algorithm

19. Computational efficiency (dynamic MRFs) Essentially, Fast-PD works along 2 different “axes”  reduces augmentations across different iterations of the same MRF  reduces augmentations across different MRFs Handles general (multi-label) dynamic MRFs Time per frame for SRI-tree stereo sequence

20. primal-dual framework Handles wide class of MRFs Approximately optimal solutions Theoretical guarantees AND tight certificates per instance Significant speed-up for static MRFs Significant speed-up for dynamic MRFs - New theorems - New insights into existing techniques - New view on MRFs Take-home message: try to take advantage of duality, whenever you can Thank you!

21. Papers and presentations available at: http://www.csd.uoc.gr/~komod/