3D Adaptive Structural Convolution Network for Domain-Invariant Point Cloud Recognition

TL;DR

The 3D Adaptive Structural Convolution Network (3D-ASCN) is introduced, a cutting-edge framework for 3D point cloud recognition that combines 3D convolution kernels, a structural tree structure, and adaptive neighborhood sampling for effective geometric feature extraction.

Abstract

Adapting deep learning networks for point cloud data recognition in self-driving vehicles faces challenges due to the variability in datasets and sensor technologies, emphasizing the need for adaptive techniques to maintain accuracy across different conditions. In this paper, we introduce the 3D Adaptive Structural Convolution Network (3D-ASCN), a cutting-edge framework for 3D point cloud recognition. It combines 3D convolution kernels, a structural tree structure, and adaptive neighborhood sampling for effective geometric feature extraction. This method obtains domain-invariant features and demonstrates robust, adaptable performance on a variety of point cloud datasets, ensuring compatibility across diverse sensor configurations without the need for parameter adjustments. This highlights its potential to significantly enhance the reliability and efficiency of self-driving vehicle technology.

Figures (3)

• Figure 1: Our proposed convolution concept and its target for different point cloud domain. (a) illustrates point cloud visualizations in different source domains. (b) shows receptive field, direction-based kernel, and distance-based kernel.
• Figure 2: Adaptive sampling and structural convolution operation: All points select their optimal neighborhoods based on k-nearest neighborhood selection and by minimizing Shannon entropy. Then, structural convolution operations are performed to extract neighborhood features of all points based on cosine similarity and Euclidean distance. Features based on cosine similarity and Euclidean distance are concatenated and passed through a multi-layer perceptron. These features are refined as they are used as inputs for this network. Finally, we can extract a global feature that includes structural information of a point cloud by aggregating all local features.
• Figure 3: Illustration of our classification architecture: Str-Conv blocks perform structural convolution. Graph Max-Pooling blocks perform channel-wise max-pooling from the features and then randomly sample a subset following the sampling rate $r$. During this process, the Adaptive neighborhood sampling block finds the optimal neighborhoods and passes them to the Str-Conv block.