NGTTA: Non-parametric Geometry-driven Test Time Adaption for 3D Point Cloud Segmentation
Keywords: Point Cloud Segmentation, Test Time Adaption
Abstract: Previous Test Time Adaption (TTA) methods usually suffer from training collapse when they are transferred to complex 3D scenes for point cloud segmentation due to the significant domain gap between the source and target data.
To solve this issue, we propose NGTTA, a stable test time adaption method guided by non-parametric geometric features. In NGTTA, we leverage the distribution of non-parametric geometric features on target data as an “intermediate domain” to reduce the domain gap and guide the stable learning of the source model on target data.
Specifically, we use the source domain model and a non-parametric geometric model to extract the embedding features and geometric features of the point cloud, respectively. Then, a category-balance sampler is designed to filter easy samples and hard samples in the input data to address the class imbalance issue in semantic segmentation. Inspired by previous work, we use easy samples for entropy minimization loss and pseudo-label prediction to fine-tune the source domain model. The difference is that we refine the pseudo labels not only by considering the soft voting among their nearest neighbors in the model embedding feature space but also in the geometric space, which can prevent the accumulation of errors caused by model feature shifts. Furthermore, we believe that hard samples can effectively represent the distribution differences between the source domain and the target domain. Therefore, we propose to distill the geometric features of hard samples into the source domain model in the early stages of training to quickly converge to an "intermediate domain" that is similar to the target domain. By taking advantage of the ability of the non-parametric geometric feature to represent the underlying manifolds of the target data, our method efficiently reduces the difficulty of the domain adaption.
We conduct the main experiments on the more challenge \textbf{\textit{sim-to-real}} benchmark about synthetic dataset 3DFRONT and the real-world datasets ScanNet and S3DIS for 3D segmentation task. Results show that our method can efficiently improve the mIOU by over \textbf{3\%} on 3DFRONT$\rightarrow$ ScanNet and \textbf{7\%} on 3DFRONT$\rightarrow$ S3DIS.
Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning
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.
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: 3379
Loading