Fast oriented bounding box optimization on the rotation group SO(3, R)

Transcription

1 Fast oriented bounding box optimization on the rotation group SO(3, R) Chia-Tche Chang 1, Bastien Gorissen 2,3 and Samuel Melchior 1,2 1 Department of Mathematical Engineering (INMA), Université catholique de Louvain 2 Department of Mechanical Engineering (MEMA), Université catholique de Louvain 3 CENAERO November 30th, 2012 C.-T. Chang et al. Fast oriented bounding box optimization on the rotation group SO(3, R) 1

2 The problem: minimum-volume OBB Coming up next... The problem: minimum-volume OBB Exact methods in 2D and 3D Classical approaches for the 3D case Our goal Bringing optimization into the game How to solve an optimization problem? Results Conclusion C.-T. Chang et al. Fast oriented bounding box optimization on the rotation group SO(3, R) 2

3 The problem: minimum-volume OBB The problem Given a set of n points X in 3D, find the minimum-volume arbitrarily oriented bounding box enclosing X. Collision detection, intersection tests, object representation, data approximation... (BV trees... ) C.-T. Chang et al. Fast oriented bounding box optimization on the rotation group SO(3, R) 3

4 The problem: minimum-volume OBB Exact methods in 2D and 3D In 2D: the rotating calipers method A minimum-area rectangle circumscribing a convex polygon has at least one side flush with an edge of the polygon. C.-T. Chang et al. Fast oriented bounding box optimization on the rotation group SO(3, R) 4

5 The problem: minimum-volume OBB Exact methods in 2D and 3D In 2D: the rotating calipers method A minimum-area rectangle circumscribing a convex polygon has at least one side flush with an edge of the polygon. Compute the convex hull: O(n log n) conv(x ) C.-T. Chang et al. Fast oriented bounding box optimization on the rotation group SO(3, R) 4

6 The problem: minimum-volume OBB Exact methods in 2D and 3D In 2D: the rotating calipers method A minimum-area rectangle circumscribing a convex polygon has at least one side flush with an edge of the polygon. Compute the convex hull: O(n log n) Loop on all edges: O(n) easy and efficient conv(x ) C.-T. Chang et al. Fast oriented bounding box optimization on the rotation group SO(3, R) 4

7 The problem: minimum-volume OBB Exact methods in 2D and 3D In 3D: generalization of the rotating calipers? A minimum-volume box circumscribing a convex polyhedron has at least one face flush with a face of the polyhedron? C.-T. Chang et al. Fast oriented bounding box optimization on the rotation group SO(3, R) 5

8 The problem: minimum-volume OBB Exact methods in 2D and 3D In 3D: generalization of the rotating calipers? A minimum-volume box circumscribing a convex polyhedron has at least one face two adjacent faces flush with a face edges of the polyhedron. [O Rourke, 1985] C.-T. Chang et al. Fast oriented bounding box optimization on the rotation group SO(3, R) 5

9 The problem: minimum-volume OBB Exact methods in 2D and 3D In 3D: generalization of the rotating calipers? A minimum-volume box circumscribing a convex polyhedron has at least one face two adjacent faces flush with a face edges of the polyhedron. [O Rourke, 1985] Problem: Loop on all pairs of edges and rotate the box while keeping edges flush O(n 3 ) time complexity... C.-T. Chang et al. Fast oriented bounding box optimization on the rotation group SO(3, R) 5

10 The problem: minimum-volume OBB Classical approaches for the 3D case In practice... O Rourke s algorithm is too slow (cubic time) use faster but inexact methods: C.-T. Chang et al. Fast oriented bounding box optimization on the rotation group SO(3, R) 6

11 The problem: minimum-volume OBB Classical approaches for the 3D case In practice... O Rourke s algorithm is too slow (cubic time) use faster but inexact methods: PCA-based methods (covariance matrix): very fast and easy to compute but may be very inaccurate C.-T. Chang et al. Fast oriented bounding box optimization on the rotation group SO(3, R) 6

12 The problem: minimum-volume OBB Classical approaches for the 3D case In practice... O Rourke s algorithm is too slow (cubic time) use faster but inexact methods: PCA-based methods (covariance matrix): very fast and easy to compute but may be very inaccurate Brute-force all orientations with a small angle increment: large computation time and/or low accuracy C.-T. Chang et al. Fast oriented bounding box optimization on the rotation group SO(3, R) 6

13 The problem: minimum-volume OBB Classical approaches for the 3D case In practice... O Rourke s algorithm is too slow (cubic time) use faster but inexact methods: PCA-based methods (covariance matrix): very fast and easy to compute but may be very inaccurate Brute-force all orientations with a small angle increment: large computation time and/or low accuracy Brute-force a well-chosen set of orientations: may sometimes have (very) good accuracy but still too slow C.-T. Chang et al. Fast oriented bounding box optimization on the rotation group SO(3, R) 6

14 The problem: minimum-volume OBB Classical approaches for the 3D case In practice... O Rourke s algorithm is too slow (cubic time) use faster but inexact methods: PCA-based methods (covariance matrix): very fast and easy to compute but may be very inaccurate Brute-force all orientations with a small angle increment: large computation time and/or low accuracy Brute-force a well-chosen set of orientations: may sometimes have (very) good accuracy but still too slow Guaranteed quality approximation methods: same problem... C.-T. Chang et al. Fast oriented bounding box optimization on the rotation group SO(3, R) 6

15 The problem: minimum-volume OBB Our goal What do we want? Goal: Very good accuracy: find an optimal OBB in (nearly?) all cases If a suboptimal solution is returned, it should be close to the best one Computational cost has to be low C.-T. Chang et al. Fast oriented bounding box optimization on the rotation group SO(3, R) 7

16 The problem: minimum-volume OBB Our goal What do we want? Goal: Very good accuracy: find an optimal OBB in (nearly?) all cases If a suboptimal solution is returned, it should be close to the best one Computational cost has to be low Our approach: iterative algorithm based on optimization methods C.-T. Chang et al. Fast oriented bounding box optimization on the rotation group SO(3, R) 7

17 Bringing optimization into the game Coming up next... The problem: minimum-volume OBB Bringing optimization into the game OBB fitting as an optimization problem Requirements Going hybrid How to solve an optimization problem? Results Conclusion C.-T. Chang et al. Fast oriented bounding box optimization on the rotation group SO(3, R) 8

18 Bringing optimization into the game OBB fitting as an optimization problem A first, direct, formulation min over size, position, orientation so that the volume of the bounding box all the points are in the box C.-T. Chang et al. Fast oriented bounding box optimization on the rotation group SO(3, R) 9

19 Bringing optimization into the game OBB fitting as an optimization problem A first, direct, formulation min over R 3, position, orientation so that the volume of the bounding box all the points are in the box = ( ξ, η, ζ ) denotes the dimensions of the OBB, C.-T. Chang et al. Fast oriented bounding box optimization on the rotation group SO(3, R) 9

20 Bringing optimization into the game OBB fitting as an optimization problem A first, direct, formulation min over R 3, position, orientation so that ξ η ζ all the points are in the box = ( ξ, η, ζ ) denotes the dimensions of the OBB, C.-T. Chang et al. Fast oriented bounding box optimization on the rotation group SO(3, R) 9

21 Bringing optimization into the game OBB fitting as an optimization problem A first, direct, formulation min over R 3, position, orientation so that ξ η ζ 1 all the rotated and 2 centered points 1 2 = ( ξ, η, ζ ) denotes the dimensions of the OBB, C.-T. Chang et al. Fast oriented bounding box optimization on the rotation group SO(3, R) 9

22 Bringing optimization into the game OBB fitting as an optimization problem A first, direct, formulation min over R 3,Ξ R 3, orientation so that ξ η ζ 1 2 all the rotated points Ξ 1 2 = ( ξ, η, ζ ) denotes the dimensions of the OBB, Ξ is the center of the OBB. C.-T. Chang et al. Fast oriented bounding box optimization on the rotation group SO(3, R) 9

23 Bringing optimization into the game OBB fitting as an optimization problem A first, direct, formulation min over R 3,Ξ R 3,R SO(3,R) so that ξ η ζ 1 2 R(all the points) Ξ 1 2 = ( ξ, η, ζ ) denotes the dimensions of the OBB, Ξ is the center of the OBB. R SO(3, R) is a rotation matrix, C.-T. Chang et al. Fast oriented bounding box optimization on the rotation group SO(3, R) 9

24 Bringing optimization into the game OBB fitting as an optimization problem A first, direct, formulation min over R 3,Ξ R 3,R SO(3,R) so that ξ η ζ 1 2 R(all the points) Ξ 1 2 = ( ξ, η, ζ ) denotes the dimensions of the OBB, Ξ is the center of the OBB. R SO(3, R) is a rotation matrix, SO(3, R) = { R R 3 3 R T R = I = RR T, det(r) = 1 }, C.-T. Chang et al. Fast oriented bounding box optimization on the rotation group SO(3, R) 9

25 Bringing optimization into the game OBB fitting as an optimization problem A first, direct, formulation min over R 3,Ξ R 3,R SO(3,R) ξ η ζ s.t. 1 2 RX i Ξ 1 2 i = 1,..., N = ( ξ, η, ζ ) denotes the dimensions of the OBB, Ξ is the center of the OBB. R SO(3, R) is a rotation matrix, SO(3, R) = { R R 3 3 R T R = I = RR T, det(r) = 1 }, X = {X i i = 1,..., N } is the considered set of points C.-T. Chang et al. Fast oriented bounding box optimization on the rotation group SO(3, R) 9

26 Bringing optimization into the game OBB fitting as an optimization problem A first, direct, formulation min R 3,Ξ R 3,R SO(3,R) ξ η ζ s.t. 1 2 RX i Ξ 1 2 i = 1,..., N = ( ξ, η, ζ ) denotes the dimensions of the OBB, Ξ is the center of the OBB. R SO(3, R) is a rotation matrix, SO(3, R) = { R R 3 3 R T R = I = RR T, det(r) = 1 }, X = {X i i = 1,..., N } is the considered set of points Smooth but constrained optimization problem C.-T. Chang et al. Fast oriented bounding box optimization on the rotation group SO(3, R) 9

27 Bringing optimization into the game OBB fitting as an optimization problem Unconstrained formulation min R SO(3,R) ( min R 3,Ξ R 3 ξ η ζ } s.t. 1 2 RX i Ξ 1 2 i = 1,..., N {{ } f (R) ) C.-T. Chang et al. Fast oriented bounding box optimization on the rotation group SO(3, R) 10

28 Bringing optimization into the game OBB fitting as an optimization problem Unconstrained formulation min R SO(3,R) ( min R 3,Ξ R 3 ξ η ζ } s.t. 1 2 RX i Ξ 1 2 i = 1,..., N {{ } f (R) ) The objective function f (R) is simply the volume of the AABB of X rotated by R C.-T. Chang et al. Fast oriented bounding box optimization on the rotation group SO(3, R) 10

29 Bringing optimization into the game OBB fitting as an optimization problem Unconstrained formulation min R SO(3,R) ( min R 3,Ξ R 3 ξ η ζ } s.t. 1 2 RX i Ξ 1 2 i = 1,..., N {{ } f (R) ) The objective function f (R) is simply the volume of the AABB of X rotated by R Unconstrained but non-differentiable optimization problem C.-T. Chang et al. Fast oriented bounding box optimization on the rotation group SO(3, R) 10

30 Bringing optimization into the game Requirements Solving this problem requires... C.-T. Chang et al. Fast oriented bounding box optimization on the rotation group SO(3, R) 11

31 Bringing optimization into the game Requirements Solving this problem requires a derivative-free method f (R) is not differentiable everywhere... C.-T. Chang et al. Fast oriented bounding box optimization on the rotation group SO(3, R) 11

32 Bringing optimization into the game Requirements Solving this problem requires a derivative-free method... a global search technique f (R) has many local minima... C.-T. Chang et al. Fast oriented bounding box optimization on the rotation group SO(3, R) 11

33 Bringing optimization into the game Requirements Solving this problem requires a derivative-free method... a global search technique... a fast convergence rate That was the point! C.-T. Chang et al. Fast oriented bounding box optimization on the rotation group SO(3, R) 11

34 Bringing optimization into the game Going hybrid Our idea: using an hybrid method 1. Use a global exploration component: genetic algorithm (GA) C.-T. Chang et al. Fast oriented bounding box optimization on the rotation group SO(3, R) 12

35 Bringing optimization into the game Going hybrid Our idea: using an hybrid method 1. Use a global exploration component: genetic algorithm (GA) 2. Speed up convergence using a local exploitation algorithm Nelder-Mead simplex algorithm (NM) C.-T. Chang et al. Fast oriented bounding box optimization on the rotation group SO(3, R) 12

36 Bringing optimization into the game Going hybrid Our idea: using an hybrid method 1. Use a global exploration component: genetic algorithm (GA) 2. Speed up convergence using a local exploitation algorithm Nelder-Mead simplex algorithm (NM) GA alone would be very slow to converge (GA more suitable for discrete search spaces) NM alone would be stuck in local minima (even with restarts) C.-T. Chang et al. Fast oriented bounding box optimization on the rotation group SO(3, R) 12

37 How to solve an optimization problem? Coming up next... The problem: minimum-volume OBB Bringing optimization into the game How to solve an optimization problem? Genetic algorithms (GA) The Nelder-Mead algorithm (NM) HYBBRID: let s mix GA and NM together! Results Conclusion C.-T. Chang et al. Fast oriented bounding box optimization on the rotation group SO(3, R) 13

38 How to solve an optimization problem? Genetic algorithms (GA) Global exploration: genetic algorithms Stochastic population-based evolutionary method (original variant proposed by Holland in the 1970s) Population-based: keep a large set of candidates at each iteration Evolutionary: generate new candidates by combining current ones depending on their performance C.-T. Chang et al. Fast oriented bounding box optimization on the rotation group SO(3, R) 14

39 How to solve an optimization problem? Genetic algorithms (GA) The general framework Start with a set of candidates (population) and a performance function (fitness function) C.-T. Chang et al. Fast oriented bounding box optimization on the rotation group SO(3, R) 15

40 How to solve an optimization problem? Genetic algorithms (GA) The general framework Start with a set of candidates (population) and a performance function (fitness function) At each generation: Selection: parents are selected depending on their fitness C.-T. Chang et al. Fast oriented bounding box optimization on the rotation group SO(3, R) 15

41 How to solve an optimization problem? Genetic algorithms (GA) The general framework Start with a set of candidates (population) and a performance function (fitness function) At each generation: Selection: parents are selected depending on their fitness Crossover: selected parents produce offsprings C.-T. Chang et al. Fast oriented bounding box optimization on the rotation group SO(3, R) 15

42 How to solve an optimization problem? Genetic algorithms (GA) The general framework Start with a set of candidates (population) and a performance function (fitness function) At each generation: Selection: parents are selected depending on their fitness Crossover: selected parents produce offsprings Mutation: offsprings can be subject to mutations (random modification, gradient step, SA step,...) C.-T. Chang et al. Fast oriented bounding box optimization on the rotation group SO(3, R) 15

43 How to solve an optimization problem? The Nelder-Mead algorithm (NM) The Nelder-Mead simplex algorithm Derivative-free simplicial optimization method (original algorithm proposed by Nelder & Mead in 1965) Simplex (in R n ) = set of n + 1 affinely independent points n = 2 : triangle n = 3 : tetrahedron... C.-T. Chang et al. Fast oriented bounding box optimization on the rotation group SO(3, R) 16

44 How to solve an optimization problem? The Nelder-Mead algorithm (NM) Ideas of the algorithm (details omitted) Reflection Expansion Contraction Reduction Four main ways to move/transform the simplex depending on the performance of its vertices: affine combinations C.-T. Chang et al. Fast oriented bounding box optimization on the rotation group SO(3, R) 17

45 How to solve an optimization problem? The Nelder-Mead algorithm (NM) Nelder-Mead and the six-hump camel back... 2 Current objective = The initial simplex C.-T. Chang et al. Fast oriented bounding box optimization on the rotation group SO(3, R) 18

46 How to solve an optimization problem? The Nelder-Mead algorithm (NM) Nelder-Mead and the six-hump camel back... 2 Current objective = Iteration 1: contraction C.-T. Chang et al. Fast oriented bounding box optimization on the rotation group SO(3, R) 18

47 How to solve an optimization problem? The Nelder-Mead algorithm (NM) Nelder-Mead and the six-hump camel back... 2 Current objective = Iteration 2: reflection C.-T. Chang et al. Fast oriented bounding box optimization on the rotation group SO(3, R) 18

48 How to solve an optimization problem? The Nelder-Mead algorithm (NM) Nelder-Mead and the six-hump camel back... 2 Current objective = Iteration 3: contraction C.-T. Chang et al. Fast oriented bounding box optimization on the rotation group SO(3, R) 18

49 How to solve an optimization problem? The Nelder-Mead algorithm (NM) Nelder-Mead and the six-hump camel back... 2 Current objective = Iteration 4: contraction C.-T. Chang et al. Fast oriented bounding box optimization on the rotation group SO(3, R) 18

50 How to solve an optimization problem? The Nelder-Mead algorithm (NM) Nelder-Mead and the six-hump camel back... 2 Current objective = Iteration 5: contraction C.-T. Chang et al. Fast oriented bounding box optimization on the rotation group SO(3, R) 18

51 How to solve an optimization problem? The Nelder-Mead algorithm (NM) Nelder-Mead and the six-hump camel back... 2 Current objective = Iteration 6: reduction C.-T. Chang et al. Fast oriented bounding box optimization on the rotation group SO(3, R) 18

52 How to solve an optimization problem? The Nelder-Mead algorithm (NM) Nelder-Mead and the six-hump camel back... 2 Current objective = Iteration 7: contraction C.-T. Chang et al. Fast oriented bounding box optimization on the rotation group SO(3, R) 18

53 How to solve an optimization problem? The Nelder-Mead algorithm (NM) Nelder-Mead and the six-hump camel back... 2 Current objective = Some more iterations... C.-T. Chang et al. Fast oriented bounding box optimization on the rotation group SO(3, R) 18

54 How to solve an optimization problem? The Nelder-Mead algorithm (NM) Nelder-Mead and the six-hump camel back... 2 Current objective = Some more iterations... C.-T. Chang et al. Fast oriented bounding box optimization on the rotation group SO(3, R) 18

55 How to solve an optimization problem? The Nelder-Mead algorithm (NM) Nelder-Mead and the six-hump camel back... 2 Current objective = The final result C.-T. Chang et al. Fast oriented bounding box optimization on the rotation group SO(3, R) 18

56 How to solve an optimization problem? HYBBRID: let s mix GA and NM together! Mixing NM and GA together (simplified version) Population of M simplices (simplex = set of 4 rotation matrices) Fitness function f (R) (volume of corresponding OBB) C.-T. Chang et al. Fast oriented bounding box optimization on the rotation group SO(3, R) 19

57 How to solve an optimization problem? HYBBRID: let s mix GA and NM together! Mixing NM and GA together (simplified version) Population of M simplices (simplex = set of 4 rotation matrices) Fitness function f (R) (volume of corresponding OBB) Selection: Evaluate fitness of all simplices, keep best 50% C.-T. Chang et al. Fast oriented bounding box optimization on the rotation group SO(3, R) 19

58 How to solve an optimization problem? HYBBRID: let s mix GA and NM together! Mixing NM and GA together (simplified version) Population of M simplices (simplex = set of 4 rotation matrices) Fitness function f (R) (volume of corresponding OBB) Selection: Evaluate fitness of all simplices, keep best 50% Crossover I: Create M 2 offsprings by mixing vertices: A 1 B 1 C 1 D 1 A 2 B 2 C 2 D 2 A i1 B i2 C i3 D i4, i k {1, 2} C.-T. Chang et al. Fast oriented bounding box optimization on the rotation group SO(3, R) 19

59 How to solve an optimization problem? HYBBRID: let s mix GA and NM together! Mixing NM and GA together (simplified version) Population of M simplices (simplex = set of 4 rotation matrices) Fitness function f (R) (volume of corresponding OBB) Selection: Evaluate fitness of all simplices, keep best 50% Crossover I: Create M 2 offsprings by mixing vertices: A 1 B 1 C 1 D 1 A 2 B 2 C 2 D 2 A i1 B i2 C i3 D i4, i k {1, 2} Crossover II: Create M 2 offsprings by affinely combine vertices: A 1 B 1 C 1 D 1 A 2 B 2 C 2 D 2 A 3 B 3 C 3 D 3 with A 3 = λa 1 + (1 λ)a 2 C.-T. Chang et al. Fast oriented bounding box optimization on the rotation group SO(3, R) 19

60 How to solve an optimization problem? HYBBRID: let s mix GA and NM together! Mixing NM and GA together (simplified version) Population of M simplices (simplex = set of 4 rotation matrices) Fitness function f (R) (volume of corresponding OBB) Selection: Evaluate fitness of all simplices, keep best 50% Crossover I: Create M 2 offsprings by mixing vertices: A 1 B 1 C 1 D 1 A 2 B 2 C 2 D 2 A i1 B i2 C i3 D i4, i k {1, 2} Crossover II: Create M 2 offsprings by affinely combine vertices: A 1 B 1 C 1 D 1 A 2 B 2 C 2 D 2 A 3 B 3 C 3 D 3 with A 3 = λa 1 + (1 λ)a 2 Mutation: Apply K Nelder-Mead iterations on each offspring C.-T. Chang et al. Fast oriented bounding box optimization on the rotation group SO(3, R) 19

61 How to solve an optimization problem? HYBBRID: let s mix GA and NM together! HYBBRID Nelder-Mead algorithm Genetic algorithm on the special orthogonal group SO(3) to solve the optimal OBB problem C.-T. Chang et al. Fast oriented bounding box optimization on the rotation group SO(3, R) 20

62 How to solve an optimization problem? HYBBRID: let s mix GA and NM together! HYBBRID Nelder-Mead algorithm Genetic algorithm on the special orthogonal group SO(3) to solve the optimal OBB problem = HYbrid Bounding Box Rotation IDentification algorithm C.-T. Chang et al. Fast oriented bounding box optimization on the rotation group SO(3, R) 20

63 Results Coming up next... The problem: minimum-volume OBB Bringing optimization into the game How to solve an optimization problem? Results Behaviour of HYBBRID Comparison to other algorithms Conclusion C.-T. Chang et al. Fast oriented bounding box optimization on the rotation group SO(3, R) 21

64 Results Behaviour of HYBBRID Behavior of HYBBRID All algorithms tested on a benchmark set of 300 objects (Gamma db) Implementations done in MATLAB (built-in functions are used) C.-T. Chang et al. Fast oriented bounding box optimization on the rotation group SO(3, R) 22

65 Results Behaviour of HYBBRID Behavior of HYBBRID: yes, it works! All algorithms tested on a benchmark set of 300 objects (Gamma db) Implementations done in MATLAB (built-in functions are used) 100 Tolerance = 1e Performance [%] Population size = Population size = Population size = 40 3 Population size = Computation time [s] Error is less than in 90%+ of the cases! C.-T. Chang et al. Fast oriented bounding box optimization on the rotation group SO(3, R) 22

66 Results Behaviour of HYBBRID How does it scale? 15 Computation time [s] Number of vertices on the convex hull Experimental results show a roughly linear complexity! C.-T. Chang et al. Fast oriented bounding box optimization on the rotation group SO(3, R) 23

67 Results Comparison to other algorithms Comparison to other iterative approaches Failure rate [%] Random search (60000 trials) Relative error tolerance Trying random orientations does not work... C.-T. Chang et al. Fast oriented bounding box optimization on the rotation group SO(3, R) 24

68 Results Comparison to other algorithms Comparison to other iterative approaches Failure rate [%] fmincon, Euler angles (15 restarts) fmincon, Rotation matrix (15 restarts) Random search (60000 trials) Relative error tolerance Constrained smooth opti: success rate 40% but mainly AABBs... C.-T. Chang et al. Fast oriented bounding box optimization on the rotation group SO(3, R) 24

69 Results Comparison to other algorithms Comparison to other iterative approaches Failure rate [%] Nelder Mead (30 restarts) fmincon, Euler angles (15 restarts) fmincon, Rotation matrix (15 restarts) Random search (60000 trials) Relative error tolerance Unconstrained non-diff. opti, random initializations: much better results! C.-T. Chang et al. Fast oriented bounding box optimization on the rotation group SO(3, R) 24

70 Results Comparison to other algorithms Comparison to other iterative approaches HYBBRID (M=50, K=30) Nelder Mead (30 restarts) fmincon, Euler angles (15 restarts) fmincon, Rotation matrix (15 restarts) Random search (60000 trials) Failure rate [%] Relative error tolerance HYBBRID: combining potential solutions does improve the success rate! C.-T. Chang et al. Fast oriented bounding box optimization on the rotation group SO(3, R) 24

71 Results Comparison to other algorithms Comparison to other algorithms First, let s ignore the computational cost and look at the failure rates AABB AABB with post processing Failure rate [%] Relative error tolerance A reference point: the simple AABB C.-T. Chang et al. Fast oriented bounding box optimization on the rotation group SO(3, R) 25

72 Results Comparison to other algorithms Comparison to other algorithms First, let s ignore the computational cost and look at the failure rates AABB AABB with post processing Min PCA with post processing Continuous PCA with post processing Failure rate [%] Relative error tolerance PCA-based methods: limited accuracy C.-T. Chang et al. Fast oriented bounding box optimization on the rotation group SO(3, R) 25

73 Results Comparison to other algorithms Comparison to other algorithms First, let s ignore the computational cost and look at the failure rates AABB AABB with post processing Min PCA with post processing Continuous PCA with post processing All Pairs Korsawe (fast variant) Korsawe (slow variant) Failure rate [%] Relative error tolerance Brute-forcing on a set of orientations may be OK... if well chosen! C.-T. Chang et al. Fast oriented bounding box optimization on the rotation group SO(3, R) 25

74 Results Comparison to other algorithms Comparison to other algorithms First, let s ignore the computational cost and look at the failure rates AABB AABB with post processing Min PCA with post processing Continuous PCA with post processing All Pairs Korsawe (fast variant) Korsawe (slow variant) Barequet & Har Peled s GridSearchMinVolBbx (d=60) Failure rate [%] Relative error tolerance Guaranteed approximation algorithms: limited by computational resources C.-T. Chang et al. Fast oriented bounding box optimization on the rotation group SO(3, R) 25

75 Results Comparison to other algorithms Comparison to other algorithms First, let s ignore the computational cost and look at the failure rates AABB AABB with post processing Min PCA with post processing Continuous PCA with post processing All Pairs Korsawe (fast variant) Korsawe (slow variant) Barequet & Har Peled s GridSearchMinVolBbx (d=60) HYBBRID(M=50, K=30) Failure rate [%] Relative error tolerance HYBBRID: more accurate than these other methods C.-T. Chang et al. Fast oriented bounding box optimization on the rotation group SO(3, R) 25

76 Results Comparison to other algorithms Comparison to other algorithms What are the computation times? (Tolerance: 10 3 ) Performance [%] Failure rate [%] Computation time [s] Relative error tolerance AABB PCA Continuous PCA Brute-force on a set of orientations O Rourke s exact algorithm Guaranteed approximation HYBBRID C.-T. Chang et al. Fast oriented bounding box optimization on the rotation group SO(3, R) 26

77 Results Comparison to other algorithms Comparison to other algorithms What are the computation times? (Tolerance: 10 6 ) Performance [%] Failure rate [%] Computation time [s] Relative error tolerance AABB PCA Continuous PCA Brute-force on a set of orientations O Rourke s exact algorithm Guaranteed approximation HYBBRID C.-T. Chang et al. Fast oriented bounding box optimization on the rotation group SO(3, R) 26

78 Conclusion Coming up next... The problem: minimum-volume OBB Bringing optimization into the game How to solve an optimization problem? Results Conclusion C.-T. Chang et al. Fast oriented bounding box optimization on the rotation group SO(3, R) 27

79 Conclusion Conclusion HYBBRID: Nelder-Mead Genetic algorithm able to approximate optimal OBBs using optimization on SO(3) C.-T. Chang et al. Fast oriented bounding box optimization on the rotation group SO(3, R) 28

80 Conclusion Conclusion HYBBRID: Nelder-Mead Genetic algorithm able to approximate optimal OBBs using optimization on SO(3) More accurate and/or faster than other algorithms C.-T. Chang et al. Fast oriented bounding box optimization on the rotation group SO(3, R) 28

81 Conclusion Conclusion HYBBRID: Nelder-Mead Genetic algorithm able to approximate optimal OBBs using optimization on SO(3) More accurate and/or faster than other algorithms Still has room for improvements... C.-T. Chang et al. Fast oriented bounding box optimization on the rotation group SO(3, R) 28

82 Conclusion Conclusion HYBBRID: Nelder-Mead Genetic algorithm able to approximate optimal OBBs using optimization on SO(3) More accurate and/or faster than other algorithms Still has room for improvements... Thank you for your attention! C.-T. Chang et al. Fast oriented bounding box optimization on the rotation group SO(3, R) 28