One 、Deep Clustering for Unsupervised Learning of Visual Features

Link to the original text ：https://arxiv.org/pdf/1807.05520.pdf

         Unsupervised learning methods that do not require labels at all , It seems to be to learn good feature expression （ Of course, downstream tasks definitely need labels ）.

Network structure

         Will all Input images into convolutional neural network to get expression （ Or characteristics ）（ Set input Corresponding output ）. Then enter into two branches ： Classification branch and clustering Branch .

         Cluster branch First cluster in some way （ Such as K-means）, Get the of each picture “ Pseudo label ”, And calculate the clustering loss （ Is the parameter matrix to be learned ）：

among .

         The taxonomic Branch Then use the classification header （） Predicted results , Compare with pseudo tags , Calculate the classification loss （log-softmax）：

         To prevent trivial solutions ： When a cluster It's empty time , Randomly select a non empty cluster , The centroid with small random disturbance is used as the empty cluster New center of mass . then , Will belong to a non empty cluster The points of are reassigned to two result clusters （ and ）; Resample the input data to make the distribution uniform .

         If the downstream task is image classification , Then discard the above classification and clustering branches , Add a new category header and fine tune .

Two 、ClusterNet: Deep Hierarchical Cluster Network with Rigorously Rotation- Invariant Representation for Point Cloud Analysis

Link to the original text ：ClusterNet: Deep Hierarchical Cluster Network With Rigorously Rotation-Invariant Representation for Point Cloud Analysis | IEEE Conference Publication | IEEE Xplore

         Strictly rotation invariant （RRI） mapping Satisfy ： And the number of points before and after mapping remains unchanged . among Is the set of points （ Point cloud ）, Transform for any rotation .

        Use the following expression ：

.

among by Of K One of the nearest neighbors , Serial number is ; Respectively -norm、 Included angle before transformation 、 The minimum value of the included angle after transformation ; Transformation To the plane （ And An orthogonal plane ） Orthogonal projection transformation .

        This paper proves that ：（a） The above expression is RRI Of ;（b） This expression can reconstruct the point cloud （ That is, there is no loss of information except rotation information ）.

         Re express the above expression as

namely .

         Then a point It can also be seen as 4 dimension 、 Point cloud of points . And then use PointNet Extract the local information of the point , obtain Whitman's sign .

Network structure

among EdgeConv（ Aggregate the characteristics of neighbors to each point , No down sampling ）：

         In addition, a method for evaluating rotation invariance is also proposed benchmark： Euler axis and rotation angle are generated uniformly and randomly , Rotate each point cloud of the test set in multiple ways （ The training set does not change ）, Then evaluate on the enhanced test set .

3、 ... and 、Unsupervised Feature Learning for Point Cloud Understanding by Contrasting and Clustering Using Graph Convolutional Neural Networks

Link to the original text ：Unsupervised Feature Learning for Point Cloud Understanding by Contrasting and Clustering Using Graph Convolutional Neural Networks | IEEE Conference Publication | IEEE Xplore

         You don't need any tags to learn the semantic features of point cloud .

ContrastNet： Divide all point clouds into two parts to get a partial set , Then two parts are randomly selected from the part set , Make a second classification judgment , That is the two one. part Whether it belongs to the same point cloud . Using cross entropy loss .

Clustering： Use K-means++ The scheme clusters the features of all point clouds , Get a fake tag .

ClusterNet： Same as the first article in this article .

         Clustering losses are as follows ：

        ClusterNet Predict the false label of the point cloud , The corresponding losses are as follows ：

summary （ Personal understanding ）

         The current deep clustering algorithm combines neural network and clustering model , The total loss is equal to the sum of network loss and clustering loss . To have clustering loss requires the objective function of clustering method to be optimized （ Such as K-means、 Spectral clustering 、 Gaussian mixture model, etc , These methods often need to specify the number of clusters ）.