WAPooling: An adaptive plug-and-play module for feature aggregation in point cloud classification networks

Deep learning methods for classification have achieved significant advancements in processing 3D point clouds. A fundamental aspect of deep learning networks is how to best aggregate features into a global representation of the point cloud. While many existing networks rely on the traditional max-pooling for feature aggregation due to its efficiency and permutation-invariance, max-pooling has some limitations. It is a fixed operation, lacking learnability and adaptability, which restricts the network’s ability to capture the most informative global feature representation. This limitation motivates for developing a new feature aggregation method that generates a richer and more diverse feature representation. Therefore, this paper introduces a novel pooling operation, the Weighted Activation Pooling (WAP) module. WAP adds adaptability to the pooling operation by using learnable weights to dynamically adjust the importance of the pooled features. Features are further transformed through different activation functions, allowing the network to learn complex patterns and relations in the data. The WAP module is a plug-and-play module that can replace the traditional pooling operation in existing networks, with minimal computational overhead. Moreover, this paper introduces the AdaptNet classification network, where the proposed WAP module is used to obtain global features. Extensive experiments are conducted to evaluate AdaptNet using both real-world data and the ModelNet40 dataset. Results show that AdaptNet outperforms other networks, achieving higher overall accuracy on both datasets. Moreover, WAP is integrated into existing classification networks, and experiments using real-world data show an increased performance of all tested networks compared to using their original pooling strategy.