PointCloud_KNN

Using Deep Neural Net PointNet to get pointcloud's vector and find K Nearest Neighbours (projection-KD tree,JL lemma))

Data

The retrieval pipeline was implemented using provided pointclouds which correspond to shapes of the following categories: bathtub, bed, chair, desk, dresser, monitor, night stand, sofa, table, toilet. The points per provided shape were 1024 and each point was placed to 3-dimensional space.

Training

I used the provided train set to train the neural network with 15 epochs. The results can be found inside the ”log” directory. After the training, I used the test set to evaluate the results and extracted the desirable vectors that describe each provided shape through evaluation on both test and train set. These vectors were extracted and saved to .npy files right after the max-pooling step of the classification and, thus, giving us 1024-dimensional points per shape.

Nearest Neighbours

All of the vectors were projected to a lower-dimensional Euclidean space (40-dimensional space in this example) following the JohnsonLindenstrauss lemma. After that the vectrors from the training set were used to construct a kd-tree whose leaves hold the indices corresponding to the original pointcloud indices. The vectors from the test set formed the queries imposed to the kd-tree. The results were the k neighbouring projected vectors from the training set for k=10,20,50,100,200,300.

Evaluation

Once the nearest neighbours were found, I computed the F1 score of the results based on caomparison between the labels of the neighbours and the label of the query. The mean values per class (10) and the total mean scores were also computed for each of the values of k mentioned above. An example plot of this visualisation can be found inside the main folder under the name figure 1.png

References [1] https://www.tensorflow.org [2] https://docs.scipy.org/doc/scipy-0.14.0/reference/generated/scipy.spatial.KDTree.html [3] https://arxiv.org/pdf/1612.00593.pdf 1[4] https://github.com/charlesq34/pointnet [5] http://scikit-learn.org/stable/modules/random projection.html 2