Research map / Point clouds / Shape and surface completion
Snowflake Point Deconvolution for Point Cloud Completion and Generation with Skip-Transformer
Xiang Peng, Xin Wen, Yu-Shen Liu, Yan‐Pei Cao, Pengfei Wan, Wen Zheng, Zhizhong Han
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| remove whole skip-transformer (no feature connection between SPD layers) (ablation) | SnowflakeNet · Completion3D 4 categories | CD x10^4: 9.39 -> 8.48 | Improved |
| remove transformer mechanism from skip-transformer, add features directly (ablation) | SnowflakeNet · Completion3D 4 categories | CD x10^4: 9.30 -> 8.48 | Improved |
| use skip-transformer instead of self-attention (ablation) | SnowflakeNet · Completion3D 4 categories | CD x10^4: 8.89 -> 8.48 | Improved |
| use point-wise splitting instead of folding-based feature expansion (ablation) | SnowflakeNet · Completion3D 4 categories | CD x10^4: 8.80 -> 8.48 | Improved |
| add preservation loss (ablation) | SnowflakeNet · Completion3D 4 categories | CD x10^4: 8.50 -> 8.48 | Improved |
| use SPD decoder on PCN encoder instead of PCN decoder (ablation) | PCN · Completion3D 4 categories | CD x10^4: 13.30 -> 8.93 | Improved |
| use three splitting steps instead of baseline two-stage design (ablation) | SnowflakeNet · Completion3D 4 categories | CD x10^4: 8.48 -> 8.66 | Hurt |
| use one-step splitting instead of baseline (ablation) | SnowflakeNet · Completion3D 4 categories | CD x10^4: 8.48 -> 9.53 | Hurt |
| increase k-NN neighbors from 4 to 32 in skip-transformer (ablation) | SnowflakeNet · Completion3D 4 categories | CD x10^4: 8.69 -> 8.37 | Improved |
| train with 0% Gaussian noise augmentation (ablation) | SnowflakeNet · Completion3D noise-free input | CD x10^4: 8.48 -> 8.81 | Hurt |
| increase test noise to 2.0% (ablation) | SnowflakeNet · Completion3D noisy input | CD x10^4: 8.81 -> 10.56 | Hurt |
| SnowflakeNet vs PMP-Net++ on PCN | PMP-Net++ · PCN | L1 CD x10^3: 7.56 -> 7.21 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 3 of 3 on most shared results.
- PoinTr — PoinTr: Diverse Point Cloud Completion With Geometry-Aware Transformers (ICCV 2021) · better on 5 of 7 shared results
- PMP-Net++ — PMP-Net: Point Cloud Completion by Learning Multi-Step Point Moving Paths (CVPR 2021) · better on 3 of 3 shared results
- GRNet — GRNet: Gridding Residual Network for Dense Point Cloud Completion (ECCV 2020) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- SnowflakeNet: Point Cloud Completion by Snowflake Point Deconvolution With Skip-Transformer (ICCV 2021 · similarity 0.98)
- CRA-PCN: Point Cloud Completion with Intra- and Inter-level Cross-Resolution Transformers (AAAI 2024 · similarity 0.88)
- GenPC: Zero-shot Point Cloud Completion via 3D Generative Priors (CVPR 2025 · similarity 0.88)
- Digging into Intrinsic Contextual Information for High-fidelity 3D Point Cloud Completion (AAAI 2025 · similarity 0.87)
- ComPC: Completing a 3D Point Cloud with 2D Diffusion Priors (ICLR 2025 · similarity 0.87)
- SymmCompletion: High-Fidelity and High-Consistency Point Cloud Completion with Symmetry Guidance (AAAI 2025 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
