传统点云配准局部描述子总结

Guo, Y., Bennamoun, M., Sohel, F., Lu, M., Wan, J., & Kwok, N. M. (2016). A Comprehensive Performance Evaluation of 3D Local Feature Descriptors. International Journal of Computer Vision, 66–89. https://doi.org/10.1007/s11263-015-0824-y

Contribution

1. 统计了2016之前的所有传统3D local Feature Descriptors
2. 专注于3D object recognition 方向进行实验，测试各描述符的descriptiveness和robustness

3D Local Feature Descriptors

个人感觉这里专注于对于使用histogram思想的描述符进行综述

they describe the local surface by accumulating geometric or topological measurements(拓扑测量) (e.g., point numbers) into histograms according to a specific domain (e.g., point coordinates, geometric attributes)

Spatial distribution histogram based Descriptors

根据local patch里点的空间分布属性，如点坐标来生成统计直方图

They usually start with the construction of a Local Reference Frame/Axis (LRF/A) for the keypoint, and partition the 3D support region into several bins according to the LRF/A. They then generate a histogram for the local surface by accumulating the spatial distribution measurements (e.g., number of points) in each spatial bin.
先建立LRF（LRA），然后在支持域上划分bins，基于空间分布测量方法生成直方图填充每个空间bin

SpinImage

keypoints：Cylindrical coordinate system、$\alpha -\beta$space、2D accumulator

3D Shape Context（3DSC）

Frome, A., Huber, D., Kolluri, R., Bülow, T., & Malik, J. (2004). Recognizing Objects in Range Data Using Regional Point Descriptors. In Lecture Notes in Computer Science,Computer Vision - ECCV 2004 (pp. 224–237). https://doi.org/10.1007/978-3-540-24672-5_18

keypoints：

1. the surface normal $n$as LRA
2. 3D spherical grid with the north pole of the grid being aligned with the surface normal $n$
3. The divisions are logarithmically spaced along the radial dimension and linearly along the other two dimensions

Unique Shape Context（USC）

Tombari, F., Salti, S., & Di Stefano, L. (2010). Unique shape context for 3d data description. Proceedings of the ACM Workshop on 3D Object Retrieval. Presented at the MM ’10: ACM Multimedia Conference, Firenze Italy. https://doi.org/10.1145/1877808.1877821

keypoints:

1. an LRF is constructed
2. local surface is aligned with the LRF for invariance to rigid transformations
3. other like 3DSC
4. USC imprtoves 3DSC interms of memory footprint and efficiency

Rotational Projection Statistics (RoPS)

Guo, Y., Sohel, F., Bennamoun, M., Lu, M., & Wan, J. (2013). Rotational Projection Statistics for 3D Local Surface Description and Object Recognition. International Journal of Computer Vision, 63–86. https://doi.org/10.1007/s11263-013-0627-y

1. an LRF is constructed
2. local patch rotated around the three corrdinate axes in LRF, with $d_{rops\_r}$(the number of rotations) around each axis
3. for each rotation, local patch further projected onto the axis planes(xy, yz, and xz)， then a distribution matrix is generated like SpinImage. In the end, every distribution matrix is encoded with five statistics:
   1. for each distribution matrix central moments are calculated: M11, M12, M21, M22, E. Here E is the Shannon entropy
4. concatenating all these statistics of all rotations and projections to construct RoPS desc: $3\times 3\times 5\times d_{rops\_r}$

Tri-Spin-Image (TriSI)

Guo, Y., Sohel, F., Bennamoun, M., Lu, M., & Wan, J. (2013). TriSI: A Distinctive Local Surface Descriptor for 3D Modeling and Object Recognition.

1. an LRF is construced
2. local patch is aligned with the LRF
3. generate three SpinImage using every axes in LRF as LRA
4. concatenating Tri-SpinImage to construct TriSI: $3d_{trisi}^2$where $d_{trisi}$is the number of bins along each dimension.

Other Methods

1. the 3D tensor descriptor:
   1. using LRF to align local patch for rigid transformation invariance
   2. support region is dvidied into several cubic bins to define a 3D grid, each element to the 3D Tensor is equal to the total surface area of the local patch which intersecting the grid bin that corresponds to that tensor element.
2. SpinImage Signature
3. multi-resolution SpinImage
4. asymmetry patterns shape context

Conclusion

共同点：基于统计知识，都有网格，网格内部用统计落点数（密度，网格内部包含的表面积）
无外乎两个流派：

1. SpinImage（投影，2维网格）派：基于SpinImage做各种优化：在LRF的三个轴都做SpinImage、做大量旋转增强……
   1. 评：能用上2维匹配的研究方法，速度快、但是信息损失要比后者大
2. shape context派（3D网格，带上3D空间去分析）：基于3DSC做各种优化：替换球形网格、用LRF……
   1. 评：速度慢一些，匹配计算方法有限、信息损失较小

展望：感觉基于空间分布+统计的方法在之后深度学习方案中借鉴的不多

Geometric Attribute Histogram based Descriptors

These descriptors represent the local surface by generating histograms according to the geometric attributes (e.g., normals, curvatures) of the points on the surface.

Local Surface Patch (LSP)

Chen, H., & Bhanu, B. (2004). 3D free-form object recognition in range images using local surface patches. International Conference on Pattern Recognition,International Conference on Pattern Recognition. https://doi.org/10.1109/icpr.2004.6

the LSP descriptor is a 2D histogram $d_{lsp\_s}\times d_{lsp\_a}$, whichi is formed by accumulating points in particular bins along the two dimensions: the shape index value and the cosine of the angle between the surface normals

1. for each point $q_i$in the support region for the keypoint $p$, the shape index and **the cosine **of the angle bertween the surface normal of $q_i$and the normal at the keypoint $p$are calculated.
2. accumulating is operated in particular bin of 2D histogram.

THRIFT

Flint, A., Dick, A., & Hengel, A. van den. (2007). Thrift: Local 3D Structure Recognition. 9th Biennial Conference of the Australian Pattern Recognition Society on Digital Image Computing Techniques and Applications (DICTA 2007). Presented at the 9th Biennial Conference of the Australian Pattern Recognition Society on Digital Image Computing Techniques and Applications (DICTA 2007), Glenelg, Australia. https://doi.org/10.1109/dicta.2007.4426794

a 1D histogram of the deviation angles between the surface normal at the keypoint $p$and the surface normals at the neighboring points { q i } \{q_i\} {qi​}

1. the contribution of each points$q_i$to a particular bin of the histogram is determined by two feactors:
   1. the density of point samples
   2. the distance from the neighrboring point to the keypoint
2. the dimension of THRIFT is $d_{thrift}$, where $d_{thrift}$is the number of bins of the histogram.

Point Feature Histogram (PFH)

Rusu, R. B., Blodow, N., Marton, Z. C., & Beetz, M. (2008). Aligning point cloud views using persistent feature histograms. 2008 IEEE/RSJ International Conference on Intelligent Robots and Systems. Presented at the 2008 IEEE/RSJ International Conference on Intelligent Robots and Systems, Nice. https://doi.org/10.1109/iros.2008.4650967

the dimension of PFH is $d_{pfh}^4$, where $d_{pfh}$is the number of bins along each dimension.

one local patch(k neighor points, one key point)， $\frac{k(k+1)}{2}$point-pairs

1. constructed a point-pair frame (a Darboux frame) for computing four geometric features of this point-pair
   1. Darboux frame:

2. four geometric features:

2. like previous work, each dimension is corresponding to one feature，count the number of points in each bin of PFH descriptor.
3. **the feature of the distance between any two points is excluded from the histogram of PFH **in order to improve its robustness with respect to variations in the point densities.

Fast Point Feature Histogram (FPFH)

Rusu, R. B., Blodow, N., & Beetz, M. (2009). Fast Point Feature Histograms (FPFH) for 3D registration. 2009 IEEE International Conference on Robotics and Automation. Presented at the 2009 IEEE International Conference on Robotics and Automation (ICRA), Kobe. https://doi.org/10.1109/robot.2009.5152473

1. construct SPFH(a simplified point feature histogram) for every point pair in the support region: In SPFH, the descriptor is generated by** concatenating three separate histograms along three feature dimensions. That is, one histogram is generated along each dimension.**
2. $FPFH=wightsum(SPFH(keypoint),SPFH(otherpointpairs))$demension:$3d_{fpfh}$

Signature of Histogram of Orientations (SHOT)

Salti, S., Tombari, F., & Di Stefano, L. (2014). SHOT: Unique signatures of histograms for surface and texture description. Computer Vision and Image Understanding, 251–264. https://doi.org/10.1016/j.cviu.2014.04.011

1. constructed a LRF for the keypoint $p$and the support region, then align local patch
2. the support region is divided into serveral volumes along the radial, azimuth and elevation axes
3. For each volume,** a local histogram** is generated by accumulating point counts into bins according to the angles between the normals at the neighboring points within the volume and the normal at the keypoint.
4. concatenating all the local histograms to construct the SHOT descriptor: $d_{shot\_r}\times d_{shot\_a}\times d_{shot\_e}\times d_{shot\_h}$, first $d$, second$d$, third$d$are respectively the number of divisions along the radial, zaimuth and elevation dimensions, and forth$d$is the number of bins in each local histogram

Other Methods

1. VD-LSD:

Taati, B., & Greenspan, M. (2011). Local shape descriptor selection for object recognition in range data. Computer Vision and Image Understanding, 115(5), 681–694. https://doi.org/10.1016/j.cviu.2010.11.021

  1. use an **eigenvalue decomposition which is operated on the covasriance matrix of each point** to obtain several invariant properties for that point
  2. The VD-LSD is then generated by **accumulating the neighboring points into histogram bins **according to their invariant properties

3. 类SIFT方法
   1. 2.5D SIFT: extending the classic SIFT method from 2D grey-scale images to depth images.

Lo, T.-W. R., & Siebert, J. P. (2009). Local feature extraction and matching on range images: 2.5D SIFT. Computer Vision and Image Understanding, 1235–1250. https://doi.org/10.1016/j.cviu.2009.06.005

  2. SI-SIFT: SIFT to extract descriptors from **the shape index values** rather than the depth values from the depth image.

Bayramoglu, N., & Alatan, A. A. (2010). Shape Index SIFT: Range Image Recognition Using Local Features. 2010 20th International Conference on Pattern Recognition. Presented at the 2010 20th International Conference on Pattern Recognition (ICPR), Istanbul, Turkey. https://doi.org/10.1109/icpr.2010.95

Conclusion

虽然都是统计落点数，但是其显然不同，基于空间分布的专注于local patch里每个点相对于关键点在什么位置，基于几何分布的专注于local pathc里点与点（有仅点与关键点之间，也有local patch里所有点对组合的方法）之间的一些几何属性：相对距离，法线偏角等。前者相当于对local patch离散化（3D网格，投影后2D网格），后者在几何属性维度上建立多维网格（descriptor），从直觉上说，前者更加直观，更加易懂，属于人类直觉，后者则是需要经过一层显式的分析，换句话说，前者包含的信息更完备，但不精细；后者包含的信息更加精细，但不完备。