Fuzzy Lightweight CNN for Point Cloud Object Classification based on Voxel

Point cloud object classification has gained attention from many researchers since the emergence of public dataset like ModelNet and ShapeNet, which contains full surface objects. However, in practice, objects captured using LiDAR are only partially covered in the scanned area, making such a task burdensome. Here, we proposed a solution to overcome those problems. It is a novel fuzzy convolutional inference (FuzzConv) incorporated with depthwise over-parameterization (DOConv). Instead of applying raw data, the point clouds are transformed into a 3D voxel. We utilized EfficientNet as our backbone and modified the Mobile inverted Bottleneck Convolution (MBConv) with DOConv. In the last fully connected (FC) layer, we added the FuzzConv layer as an inference before feeding the feature map to the output layer. Consequently, to validate the performance of our model, we undertake an evaluation with multiple classifications in ModelNet10, ModelNet40, and our core dataset, the point cloud of human poses. Accuracy, loss, number of parameters, loss, precision, and F1-scores are employed as performance indicators. As a result, our model achieved top performance regarding the accuracy and loss value for the primary dataset, 83 % and 0.56, for ModelNet10 88.1 % and 0.56, and ModelNet40 74.1 % and 1.15.