CVPR Findings2026
HEDA: Hyperbolic-Euclidean Dual Adaptation for Robust Real-World Point Cloud Completion
Aihua Mao, Jun Yang, Yong-Jin Liu, Ying He
Abstract
Completing point clouds from real-world data remains challenging due to the large synthetic-to-real domain gap and the absence of ground-truth supervision for real scans. Existing adaptation methods typically rely on implicit feature alignment in Euclidean space, which cannot fully capture hierarchical geometric relationships and provide strong structural constraints, thereby limiting cross-domain generalization. To address these challenges, we present HEDA (Hyperbolic-Euclidean Dual Adaptation), a framework that unifies implicit hyperbolic adaptation with explicit Euclidean feature alignment. Specifically, we employ a part-based decomposition module to segment the input into semantically consistent parts and use a shared encoder to extract part-level features. These features are regularized by a normalizing flow into a tractable latent distribution and then embedded in hyperbolic space, where hierarchical object-part relations are naturally preserved. A Hyperbolic Part-Aware Transformer and a Hyperbolic Geometric Decoder operate within this curved space to reconstruct complete shapes. For implicit adaptation, we apply a discriminator to the hyperbolic features from both synthetic and real domains, trained with adversarial loss. Complementing the implicit adaptation, a Euclidean Self-Adaptive Anchor Alignment module explicitly aligns structural anchor features between synthetic and real domains via maximum mean discrepancy, enforcing geometric consistency across domains. Extensive experiments on the USSPA and ScanObjectNN benchmarks demonstrate that our method achieves state-of-the-art performance, significantly improving both reconstruction accuracy and visual fidelity for real-world point cloud completion.
Ask Lune
Know this paper in minutes, not an afternoon.
Lune can read every page of it and the top-tier papers that cite it, and gives you its claims, its evidence and its limits, each one cited to the paper it comes from. No skimming, no AI slop.
Lineage
- SnowflakeNet: Point Cloud Completion by Snowflake Point Deconvolution with Skip-Transformer cites Morphing and Sampling Network for Dense Point Cloud Completion
- SnowflakeNet: Point Cloud Completion by Snowflake Point Deconvolution with Skip-Transformer cites Unpaired Point Cloud Completion on Real Scans using Adversarial Training
Builds on29
- Revisiting Point Cloud Classification: A New Benchmark Dataset and Classification Model on Real-World DataMikaela Angelina Uy, Quang-Hieu Pham, Binh-Son Hua, Duc Thanh Nguyen et al.ICCV 2019 · 1,003 citations
- Morphing and Sampling Network for Dense Point Cloud CompletionMinghua Liu, Lu Sheng, Sheng Yang, Jing Shao et al.AAAI 2020 · 363 citations
- SnowflakeNet: Point Cloud Completion by Snowflake Point Deconvolution with Skip-TransformerPeng Xiang, Xin Wen, Yu-Shen Liu, Yan-Pei Cao et al.ICCV 2021 · 318 citations
- Unpaired Point Cloud Completion on Real Scans using Adversarial TrainingXuelin Chen, Baoquan Chen, Niloy J. MitraICLR 2020 · 146 citations
- ASFM-Net: Asymmetrical Siamese Feature Matching Network for Point CompletionYaqi Xia, Yan Xia, Wei Li, Rui Song et al.ACM MM 2021 · 93 citations
Related papers
- DAPointMamba: Domain Adaptive Point Mamba for Point Cloud CompletionYinghui Li, Qianyu Zhou, Di Shao, Hao Yang et al.AAAI 2026 · 1 citation
- Hyperspherical Embedding for Point Cloud CompletionJunming Zhang, Haomeng Zhang, Ram Vasudevan, Matthew Johnson-RobersonCVPR 2023
- Hyperbolic-Constraint Point Cloud Reconstruction from Single RGB-D ImagesWenrui Li, Zhe Yang, Wei Han, Hengyu Man et al.AAAI 2025 · 7 citations
- Hyperbolic Chamfer Distance for Point Cloud CompletionFangzhou Lin, Yun Yue, Songlin Hou, Xuechu Yu et al.ICCV 2023 · 53 citations
- Rethinking the compositionality of point clouds through regularization in the hyperbolic spaceAntonio Montanaro, Diego Valsesia, Enrico MagliNeurIPS 2022 · 46 citations
