Fast and Robust Normal Estimation for Point Clouds with Sharp Features

Transcription

1 1/37 Fast and Robust Normal Estimation for Point Clouds with Sharp Features Alexandre Boulch & Renaud Marlet University Paris-Est, LIGM (UMR CNRS), Ecole des Ponts ParisTech Symposium on Geometry Processing 2012

2 2/37 Normal estimation Normal estimation for point clouds Our method Experiments Conclusion

3 3/37 Normal Estimation Normal estimation for point clouds Our method Experiments Conclusion

4 4/37 Data Point clouds from photogrammetry or laser acquisition: may be noisy Sensitivity to noise Robustness to noise

5 4/37 Data Point clouds from photogrammetry or laser acquisition: may be noisy may have outliers Regression Plane P

6 4/37 Data Point clouds from photogrammetry or laser acquisition: may be noisy may have outliers most often have sharp features Smoothed sharp features Preserved sharp features

7 4/37 Data Point clouds from photogrammetry or laser acquisition: may be noisy may have outliers most often have sharp features may be anisotropic Sensitivity to anisotropy Robustness to anisotropy

8 4/37 Data Point clouds from photogrammetry or laser acquisition: may be noisy may have outliers most often have sharp features may be anisotropic may be huge (more than 20 million points)

9 5/37 Normal Estimation Normal estimation for point clouds Our method Experiments Conclusion

10 6/37 Basics of the method (2D case here for readability) Let P be a point and N P be its neighborhood.

11 6/37 Basics of the method (2D case here for readability) Let P be a point and N P be its neighborhood. We consider two cases: P lies on a planar surface

12 6/37 Basics of the method (2D case here for readability) Let P be a point and N P be its neighborhood. We consider two cases: P lies on a planar surface

13 6/37 Basics of the method (2D case here for readability) Let P be a point and N P be its neighborhood. We consider two cases: P lies on a planar surface P lies next to a sharp feature

14 6/37 Basics of the method (2D case here for readability) Let P be a point and N P be its neighborhood. We consider two cases: P lies on a planar surface P lies next to a sharp feature

15 6/37 Basics of the method (2D case here for readability) Let P be a point and N P be its neighborhood. We consider two cases: P lies on a planar surface P lies next to a sharp feature If Area(N 1 ) > Area(N 2 ), picking points in N 1 N 1 is more probable than N 2 N 2, and N 1 N 2 leads to random normals.

16 Basics of the method (2D case here for readability) Main Idea Draw as many primitives as necessary to estimate the normal distribution, and then the most probable normal. P Discretize the problem Normal direction N.B. We compute the normal direction, not orientation. 7/37

17 Basics of the method (2D case here for readability) Main Idea Draw as many primitives as necessary to estimate the normal distribution, and then the most probable normal. P Discretize the problem Fill a Hough accumulator Normal direction N.B. We compute the normal direction, not orientation. 7/37

18 Basics of the method (2D case here for readability) Main Idea Draw as many primitives as necessary to estimate the normal distribution, and then the most probable normal. P Discretize the problem Fill a Hough accumulator Select the good normal Normal direction N.B. We compute the normal direction, not orientation. 7/37

19 8/37 Robust Randomized Hough Transform T, number of primitives picked after T iteration. T min, number of primitives to pick M, number of bins of the accumulator ˆp m, empirical mean of the bin m p m, theoretical mean of the bin m

20 9/37 Robust Randomized Hough Transform Global upper bound T min such that: P( max m {1,...,M} ˆp m p m δ) α From Hoeffding s inequality, for a given bin: P( ˆp m p m δ) 2 exp( 2δ 2 T min ) Considering the whole accumulator: T min 1 2M ln( 2δ2 1 α )

21 10/37 Robust Randomized Hough Transform Confidence Interval Idea: if we pick often enough the same bin, we want to stop drawing primitives. From the Central Limit Theorem, we can stop if: 1 ˆp m1 ˆp m2 2 T i.e. the confidence intervals of the most voted bins do not intersect (confidence level 95%)

22 11/37 Accumulator Our primitives are planes directions (defined by two angles). We use the accumulator of Borrmann & al (3D Research, 2011). Fast computing Bins of similar area

23 12/37 Discretization issues P The use of a discrete accumulator may be a cause of error.

24 12/37 Discretization issues P The use of a discrete accumulator may be a cause of error.

25 12/37 Discretization issues P The use of a discrete accumulator may be a cause of error. Solution Iterate the algorithm using randomly rotated accumulators.

26 13/37 Normal Selection P Normal directions obtained by rotation of the accumulator Mean over all the normals Best confidence Mean over best cluster

27 14/37 Dealing with anisotropy The robustness to anisotropy depends of the way we select the planes (triplets of points) Sensitivity to anisotropy Robustness to anisotropy

28 15/37 Random point selection among nearest neighbors Dealing with anisotropy The triplets are randomly selected among the K nearest neighbors. Fast but cannot deal with anisotropy.

29 16/37 Uniform point selection on the neighborhood ball Dealing with anisotropy Pick a point Q in the neighborhood ball Q P

30 16/37 Uniform point selection on the neighborhood ball Dealing with anisotropy Pick a point Q in the neighborhood ball Consider a small ball around Q Q P

31 16/37 Uniform point selection on the neighborhood ball Dealing with anisotropy Pick a point Q in the neighborhood ball Consider a small ball around Q Pick a point randomly in the small ball Q P

32 16/37 Uniform point selection on the neighborhood ball Dealing with anisotropy Pick a point Q in the neighborhood ball Consider a small ball around Q Pick a point randomly in the small ball Iterate to get a triplet Deals with anisotropy, but for a high computation cost.

33 17/37 Cube discretization of the neighborhood ball Dealing with anisotropy Discretize the neighborhood ball P

34 17/37 Cube discretization of the neighborhood ball Dealing with anisotropy Discretize the neighborhood ball Pick a cube P

35 17/37 Cube discretization of the neighborhood ball Dealing with anisotropy Discretize the neighborhood ball Pick a cube Pick a point randomly in this cube P

36 17/37 Cube discretization of the neighborhood ball Dealing with anisotropy Discretize the neighborhood ball Pick a cube Pick a point randomly in this cube Iterate to get a triplet Good compromise between speed and robustness to anisotropy.

37 18/37 Normal Estimation Normal estimation for point clouds Our method Experiments Conclusion

38 19/37 Methods used for comparison Regression Hoppe & al (SIGGRAPH,1992): plane fitting Cazals & Pouget (SGP, 2003): jet fitting Noise Outliers Sharp fts Anisotropy Fast Plane fitting Jet fitting

39 19/37 Methods used for comparison Regression Hoppe & al (SIGGRAPH,1992): plane fitting Cazals & Pouget (SGP, 2003): jet fitting Voronoï diagram Dey & Goswami (SCG, 2004): NormFet Noise Outliers Sharp fts Anisotropy Fast Plane fitting Jet fitting NormFet

40 19/37 Methods used for comparison Regression Hoppe & al (SIGGRAPH,1992): plane fitting Cazals & Pouget (SGP, 2003): jet fitting Voronoï diagram Dey & Goswami (SCG, 2004): NormFet Sample Consensus Models Li & al (Computer & Graphics, 2010) Plane fitting Jet fitting NormFet Sample Consensus Noise Outliers Sharp fts Anisotropy Fast

41 20/37 Precision Two error measures: Root Mean Square (RMS): RMS = 1 2 n P,ref n P,est C P C Root Mean Square with threshold (RMS_τ): RMS_τ = 1 C vp 2 P C where { v P = n P,ref n P,est if n P,ref n P,est < τ π 2 otherwise More suited for sharp features

42 21/37 Visual on error distances Same RMS, different RMS τ

43 22/37 Precision (with noise) Precision for cube uniformly sampled, depending on noise.

44 23/37 Precision (with noise and anisotropy) Precision for a corner with anisotropy, depending on noise.

45 24/37 Computation time Computation time for sphere, function of the number of points.

46 25/37 Robustness to outliers Noisy model (0.2%) + 100% of outliers.

47 26/37 Robustness to outliers Noisy model (0.2%) + 200% of outliers.

48 Robustness to anisotropy 27/37

49 Preservation of sharp features 28/37

50 29/37 Robustness to natural noise, outliers and anisotropy Point cloud created by photogrammetry.

51 30/37 Normal Estimation Normal estimation for point clouds Our method Experiments Conclusion

52 31/37 Conclusion Plane fitting Jet fitting NormFet Sample Consensus Our method Noise Outliers Sharp fts Anisotropy Fast Compared to state-of-the-art methods that preserve sharp features, our normal estimator is: at least as precise at least as robust to noise and outliers almost 10x faster robust to anisotropy

53 32/37 Code available Web site Two versions under GPL license: for Point Cloud Library ( for CGAL (

54 33/37 Computation time T min =700 T min =300 n rot =5 n rot =2 w/o with w/o with Model (# vertices) interv. interv. interv. interv. Armadillo (173k) 21 s 20 s 3 s 3 s Dragon (438k) 55 s 51 s 8 s 7 s Buddha (543k) s 10 s Circ. Box (701k) s 12 s Omotondo (998k) s 10 s Statuette (5M) Room (6.6M) Lucy (14M)

55 34/37 Parameters K or r: number of neighbors or neighborhood radius, T min : number of primitives to explore, n φ : parameter defining the number of bins, n rot : number of accumulator rotations, c: presampling or discretization factor (anisotropy only), a cluster : tolerance angle (mean over best cluster only).

56 Efficiency 35/37

57 36/37 Influence of the neighborhood size K = 20 K = 200 K = 400

58 37/37 Precision (with noise) Precision for cube uniformly sampled, depending on noise.