Dynamic Taylor Convolutional Neural Network for Few-Shot Point Cloud Semantic Segmentation
Keywords: Point Cloud Semantic Segmentation; Few-Shot Learning; Taylor Series; Dynamic Convolution; Prototype Refinement
Abstract: Few-shot Point cloud semantic segmentation remains a challenge in the field of computer vision due to the limitations of the pre-training learning paradigm and insufficient local geometric structure representation. To address this issue, we propose a novel pre-training-free Dynamic Taylor Convolutional Neural Network, called DyTaylorCNN ingeniously, which combines the potential of the Taylor series in local structure representation with the flexibility and adaptability of dynamic convolutions. The core of DyTaylorCNN lies in two innovative components: the Dynamic Taylor Convolution (DyTaylorConv) and the Interactive Prototype Refinement (IPR) Module. Inspired by the Taylor series and dynamic convolution, DyTaylorConv performs local structure fitting by collaborating between the Low-order Convolution (LoConv) and the Dynamic High-order Convolution (DyHiConv). LoConv is designed based on position encoding, focusing on extracting the basic geometric information of point clouds, while DyHiConv adaptively models complex local geometric features by learning spatial priors to generate dynamic weights. Moreover, the IPR Module effectively reduces the domain distribution gap by learning fine-grained prototype features, further enhancing the model's generalization capability. Experimental results on multiple benchmark datasets demonstrate that the proposed DyTaylorCNN significantly outperforms current state-of-the-art methods.
Primary Area: applications to computer vision, audio, language, and other modalities
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 11034
Loading