Learning Key Features Transformer Network for Point Cloud Processing

Due to the unordered and irregular nature of point cloud data, it is challenging for neural networks to learn from it. Attention mechanisms have shown promising results in point cloud processing. It is also inherently permutation-invariant when dealing with a set of points, which makes it ideal for point cloud learning. In this paper, an attention-based encoder-decoder architecture called KFT-Net (Key Features Transformer) is designed for point cloud classification and segmentation tasks. KFT-Net has improved upon previous methods by capturing long-range contextual information, preserving important attention scores, and utilizing convolutional neural networks to extract local features. Additionally, to enhance the computational efficiency, we introduce the Top-k operation into the attention mechanism and utilize the average pooling operation to improve attention score calculation and feature extraction efficiency. Extensive experiments validate the effectiveness of the KFT-Net, demonstrating impressive performance in point cloud classification and segmentation tasks.