3D Shape Retrieval "Scan to search"

Demo

Video Demo

Input

Output

Main algorithm - Spherical Harmonics

Spherical Harmonics Basic Idea:
(1)Fouries transform can describe the distribution of a signal, and how the signal changes along time/space domain
(2)Spherical Harmonics can describe the distribution of a "signal" in space(sphere with certain radius), and how the "signal" changes along space domain

Structure

1. Pre-processing:
   (1)Normalize
   -make the center of mass of the model be at the point (R,R,R)
   -scale so that the average distance from vertices to the center of mass is R/2
   -(scale invariance and against outliers)
   (2)Denoise
   -use 3D bilateral filter to denoise
   -(scanned noise invariance)
   (3)Rasterize
   -rasterize in to a 2R2R2R voxel grid (normally choose R to be ~32)
   -(provide adequate granularity for discriminating shapes while filtering out high-freq noise)

2. Descriptor:
   Spherical harmonics descriptors(SH)
   Distance histogram descriptors(DH)

3. Retrieval: SH and DH database
   Retrieval with same weight (with 0.5 on SH and DH respectively)

Rasterization

O(n) solution:

/* fill voxel grid and get rasterized points in O(n) */
bitmap = new bitmap(2R*2R*2R)               //temporary bitmap
for each voxel to be filled
    if(bitmap(voxel.x,voxel.y,voxel.z))==0  //has not been filled or registered
        bitmap(voxel.x,voxel.y,voxel.z)=1
        grid_point.push_back(voxel)
    end
end
delete bitmap;

Spherical harmonics descriptor

Pseudo code for spherical harmonics:

//SH is spherical harmonics descriptor 
sort vertex according to radius
for each frequency l
    for each rasterized vertex in one radius range r 
        calculate F_lr = F(theta,phi)
    end
    a_ml = sum of F_lr in one radius ragion
    SH(l,r) += abs(a_ml)^2    
end
SH = sqrt(SH)

*where F_lr is the following equation, l is m in the equation

Function: double gsl_sf_legendre_sphPlm (int l, int m, double x)
Function: int gsl_sf_legendre_sphPlm_e (int l, int m, double x, gsl_sf_result * result)
These routines compute the normalized associated Legendre polynomial \sqrt{(2l+1)/(4\pi)} \sqrt{(l-m)!/(l+m)!} P_l^m(x) suitable for use in spherical harmonics. The parameters must satisfy m >= 0, l >= m, |x| <= 1. Theses routines avoid the overflows that occur for the standard normalization of P_l^m(x).

The Legendre polynomial P(n,x) can be defined by:

P(0,x) = 1
P(1,x) = x
P(n,x) = (2*n-1)/n * x * P(n-1,x) - (n-1)/n * P(n-2,x)

Further improvement

1.database clustering

2.rasterization algorithm updates

3.Spherical Harmonics transform speed up
(1)divide and concour
http://www.ams.org/journals/mcom/2002-71-238/S0025-5718-01-01386-2/
(2)FFT to fourier then SH:
http://connection.ebscohost.com/c/articles/67655125/3d-objects-retrieval-using-spherical-harmonics-feature-vector-method
(3)O(n) solution for spherical harmonics:
http://liris.cnrs.fr/Documents/Liris-2276.pdf

4.New descriptors

Reference:

1. Spherical harmonics and Legendre polynomials ,involving solution when m is negative:
   http://blog.sciencenet.cn/blog-548663-715825.html
2. Rodrigues' formula: equation dn/dxn = (d/dx)n:
   http://wenku.baidu.com/view/72c151ef102de2bd960588f2
3. Rotation Invariant Spherical Harmonic of 3D Shape:
   http://www.chenkuantong.com/?p=1210
4. An explaination of a_mn component(arbitrary constants?):
   https://www.ngs.noaa.gov/PUBS_LIB/Geodesy4Layman/TR80003F.HTM
5. Spherical harmonics library:
   http://www.cs.dartmouth.edu/~geelong/sphere/
6. Spherical Harmonics Visual Representation:
   http://afj-phd.blogspot.co.uk/2008/11/spherical-harmonics-visual.html
7. GSL library to calculate Spherical harmonics:
   https://www.gnu.org/software/gsl/manual/html_node/Associated-Legendre-Polynomials-and-Spherical-Harmonics.html
8. Shape Descriptors from John Hopkins
   http://www.cs.jhu.edu/~misha/Code/ShapeSPH/ShapeDescriptor/