Introduction

CVPR22-IMVC-CBG

This repo is a MATLAB implementaion of Highly-efficient Incomplete Large-scale Multi-view Clustering with Consensus Bipartite Graph in CVPR2022

Understanding

We think introducing anchor learning into IMVC can benefit large-scale tasks. For better understading, we strongly recommend reading Notes on implementing large-scale IMVC with anchor graphs.

Algorithm steps

Step1: Generating partial incomplete multi-view datasets with incompelte ratio from 0.1 to 0.9

Using the 'Incomplete/randomlyGeneratePartialData.m' provided by Professor Chang Tang in 'High-Order Correlation Preserved Incomplete Multi-View Subspace Clustering' published in IEEE TIP2022.

Step2: run run.m

Step3: Another improved version is in the /Improved version/ files.

Time

To further speed up the algorithm, we can use parfor in Matlab for Parallel Computing while the first time run will cost some time. For large-scale tasks, it is time-saving.

Parallel work code

Scalable Partial Multi-view Clustering with Consistent Anchor Graph

We found initialation important for large-scale IMVC tasks. We are trying to accomplish a deep neural network for new work. Advice is welcome.

Given example

In 'Incomplete' files, we provide the incomplete datasets for Caltech101-7/20/BDGP.

Randomness

The results may be slightly different with the $k$-means. (We report 20 runs and report the avearge)

Implementation details

Provided key functions for future work:

EprojSimplex.m:

funtion to sovle anchor graph $\mathbf{Z}$, provided by Weiran Wang -- Projection onto the capped simplex

other optimization:

In machine learning and computer vision community, the used optimization is called Orthogonal Procrustes Analysis which has been well studied in literature.

---

Notice: There is no need for constructing matrix A $\in \mathbb{R}^{n \times n_i}$ ($n_i$ donotes the number of existing samples) as mentioned in the paper.(constructA.m) Only findindex.m is used in large-scale incomplete multi-view clustering.

Connection

Thanks. Any problem can contact Siwei Wang(wangsiwei13@nudt.edu.cn).