【论文阅读】23-Robust Feature Matching Using Spatial Clustering With Heavy Outliers
【论文阅读】22-Robust Feature Matching Using Spatial Clustering With Heavy Outliers
- 0、basic info
- 1、method essence
- 2、step1、putative matches
- 3、step2、feature matching –2—spatial sample
- 4、step3、space clustering
-
- 4.1、 RFM-SCAN
- 4.2、 input参数设置
- 4.3、K-dist
-
- 4.3.1、K
- 4.3.2、距离测度
- 4.4、自适应参数
-
- 4.4.1、邻域距离
- 4.4.2、MinPt = k
- 5、改进后：DBSCAN
-
- 5.1、Step3：基于上文，计算自适应参数邻域距离、MinPt
- 5.2、Step4：计算core sample
- 5.3、Step5：计算density-based inlier samples
- 5.4、Step6：Label outlierID
- 6、实验结果
0、basic info
Ma J , Jiang X , Jiang J , et al. Robust Feature Matching Using Spatial Clustering With Heavy Outliers[J]. IEEE Transactions on Image Processing, 2019, PP(99):1-1.
1、method essence
- 原理： Motion consistency – small local neighborhood
- 具体细节：
将feature matching 转化为spatial sample，基于noise sample，进行聚类；非clusters的sample，视作outlier
2、step1、putative matches
略
3、step2、feature matching –2—spatial sample
- putative matches
- motion vector
4、step3、space clustering
4.1、 RFM-SCAN
多次迭代，改进DBSCAN:
4.2、 input参数设置
通过实验（不同参数值；F-score为指标；多种数据集：30 randomly chosen image
pairs involving different types of transformations），得到全局最优参数
4.3、K-dist
4.3.1、K
N：sample number
4.3.2、距离测度
enhance the motion consistence among neighboring feature points
distance measurement function：Gaussian kernel distance //Euclidean distance(本文)
4.4、自适应参数
通过实验说明：predefined parameter 效果不好
衡量指标：recall、precision、F-score
补充知识：DBSCAN启发式参数设置
4.4.1、邻域距离
4.4.2、MinPt = k
5、改进后：DBSCAN
5.1、Step3：基于上文，计算自适应参数邻域距离、MinPt
5.2、Step4：计算core sample
此方法，对cluster border inliers而言，不稳定，容易造成误判
解决办法：density-based inlier(见下文5.3);
5.3、Step5：计算density-based inlier samples
5.4、Step6：Label outlierID
6、实验结果
- 代码语言
- 设备信息
- 结果显示：show at most 200 feature matches
- Ground truth：manually checking
- Robust to image rotation && scale (small neighborhood内的motion consistency仍成立)
- State-of-the-art feature matching methods：
[18] X. Li and Z. Hu, “Rejecting mismatches by correspondence function,”
Int. J. Comput. Vis., vol. 89, no. 1, pp. 1–17, 2010.
[22] H. Liu and S. Yan, “Common visual pattern discovery via spatially
coherent correspondence,” in Proc. IEEE Conf. Comput. Vis. Pattern
Recognit., Jun. 2010, pp. 1609–1616.
[11] J. Ma, J. Wu, J. Zhao, J. Jiang, H. Zhou, and Q. Z. Sheng, “Nonrigid
point set registration with robust transformation learning under manifold
regularization,” IEEE Trans. Neural Netw. Learn. Syst., to be published.
doi: 10.1109/TNNLS.2018.2872528.
[1] J. Ma, J. Zhao, J. Jiang, H. Zhou, and X. Guo, “Locality preserving
matching,” Int. J. Comput. Vis., vol. 127, no. 5, pp. 512–531, 2019.
[5] K. M. Yi, E. Trulls, Y. Ono, V. Lepetit, M. Salzmann, and P. Fua,
“Learning to find good correspondences,” in Proc. IEEE Conf. Comput.
Vis. Pattern Recognit., Jun. 2018, pp. 2666–2674.
- LFGC：high precision but low recall
7. 数据集：
8. 定量分析：precision – recall running time
9. RFM-SCAN（本文方法）：outlier incresing
原因：
10. 应用：co-segmentation
- Initial segmentation：super-pixel segmentation method – linear spectral clustering
- Matches number — background // foreground
适用情况 ：前背景matches number 相差较大
- 定量分析分割效果：
指标：
- State-of-art methods：
[47] K. R. Jerripothula, J. Cai, and J. Yuan, “Image co-segmentation
via saliency co-fusion,” IEEE Trans. Multimedia, vol. 18, no. 9,
pp. 1896–1909, Sep. 2016.
[48] X. Dong, J. Shen, L. Shao, and M.-H. Yang, “Interactive cosegmentation
using global and local energy optimization,” IEEE Trans. Image
Process., vol. 24, no. 11, pp. 3966–3977, Nov. 2015.
[49] W. Wang and J. Shen, “Higher-order image co-segmentation,” IEEE
Trans. Multimedia, vol. 18, no. 6, pp. 1011–1021, Jun. 2016.