Research map / 3D vision / 3D generation
ReassembleNet: Learnable Keypoints and Diffusion for 2D Fresco Reconstruction
Adeela Islam, Stefano Fiorini, Stuart James, Pietro Morerio, Alessio Del Bue
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 |
|---|---|---|---|
| add learnable keypoint selector instead of farthest point sampling (ablation) | ReassembleNet (fine-tuned) · RePAIR 2D | RMSE rotation (deg): 42.98 -> 32.91 | Improved |
| add learnable keypoint selector instead of farthest point sampling (ablation) | ReassembleNet (no fine-tuning) · RePAIR 2D | RMSE rotation (deg): 55.01 -> 47.61 rotation RMSE; Qpos 0.35 -> 0.27 | Mixed |
| add trainable instead of frozen keypoint selector (ablation) | ReassembleNet (fine-tuned) · RePAIR 2D | RMSE translation (mm): 18.41 -> 17.18 | Improved |
| add fine-tuning on RePAIR after semi-synthetic pretraining (ablation) | ReassembleNet w/ Learnable KP selection · RePAIR 2D | RMSE rotation (deg): 47.61 -> 32.91 | Improved |
| add fine-tuning after pretraining (ablation) | ReassembleNet w/ Learnable KP selection · RePAIR 2D | Qpos: 0.27 -> 0.21 | Hurt |
| increase number of keypoints k from 3 to 20 (ablation) | ReassembleNet · RePAIR 2D | RMSE rotation (deg): 103.26 -> 47.61 | Improved |
| increase number of keypoints k from 3 to 20 (ablation) | ReassembleNet · RePAIR 2D | GPU memory (MB): 22630 -> 36701 | Hurt |
| use erosion and random jitter in semi-synthetic fragment generation (ablation) | ReassembleNet pretraining · RePAIR 2D | RMSE rotation (deg): 56.54 -> 32.91 | Improved |
| use erosion and random jitter in semi-synthetic fragment generation (ablation) | ReassembleNet pretraining · RePAIR 2D | RMSE translation (mm): 26.42 -> 17.18 | Improved |
| add geometric, local and global texture features to learnable-selector model (no transfer) (ablation) | ReassembleNet Frozen KP selection · RePAIR 2D | RMSE rotation (deg): 62.45 -> 51.96 | Improved |
| add all features and transfer learning (ablation) | ReassembleNet w/ Learnable KP selection · RePAIR 2D | RMSE rotation (deg): 47.61 -> 32.91 | Improved |
| treat pieces as keypoint sets instead of square patches | ReassembleNet vs DiffAssemble · RePAIR 2D (fine-tuning) | RMSE rotation (deg): 123.42 -> 32.91 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- DiffAssemble — DiffAssemble: A Unified Graph-Diffusion Model for 2D and 3D Reassembly (CVPR 2024) · better on 5 of 5 shared results
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
- DiffAssemble: A Unified Graph-Diffusion Model for 2D and 3D Reassembly (CVPR 2024 · similarity 0.88)
- Image Reassembly Combining Deep Learning and Shortest Path Problem (ECCV 2018 · similarity 0.87)
- PHFormer: Multi-Fragment Assembly Using Proxy-Level Hybrid Transformer (AAAI 2024 · similarity 0.85)
- E-M3RF: An Equivariant Multimodal 3D Re-assembly Framework (ECCV 2026 · similarity 0.85)
- UniSDF: Unifying Neural Representations for High-Fidelity 3D Reconstruction of Complex Scenes with Reflections (NeurIPS 2024 · similarity 0.82)
- Learning Partonomic 3D Reconstruction from Image Collections (CVPR 2025 · similarity 0.82)
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.
