DCNet: exploring fine-grained vision classification for 3D point clouds

Fine-grained 3D point cloud classification is vital for shape analysis and understanding. However, due to the subtle inter-class differences and the significant intra-class variations, applying the existing point cloud network directly to fine-grained visual classification tasks may suffer overfitting and cannot achieve good performance. To address this problem, we propose a unified and robust learning framework, named dynamic confusion network (DCNet), which helps the network capture the subtle differences between samples from different sub-categories more robustly. Specifically, in the stage of feature extraction, we design a novel mutual complementary mechanism between an attention block and a dynamic sample confusion block to extract more abundant discriminative features. Furthermore, we construct robust adversarial learning between a dynamic sample confusion loss and a cross-entropy loss based on a siamese network framework to make the network learn more stable feature distributions. We conduct comprehensive experiments and show that DCNet achieves the best performance in three fine-grained categories, with relative accuracy improvements of 1.35%, 1.28%, and 2.30% on Airplane, Car, and Chair, respectively, compared to state-of-the-art point cloud methods. In addition, our approach also achieves comparable performance for the coarse-grained dataset on ModelNet40.

Automated summary

In this study, a learning framework called dynamic confusion network (DCNet) is proposed, which captures subtle differences between samples from different sub-categories more robustly in fine-grained 3D point cloud classification. The experiments show that DCNet outperforms state-of-the-art methods and achieves the best performance in three fine-grained categories.