Research map / Vision-language and multimodal / Human and embodied modeling
The Missing GAP: From Solving Square Jigsaw Puzzles to Handling Real World Archaeological Fragments
Ofir Itzhak Shahar, Gur Elkin, Ohad Ben-Shahar
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 iterative discrete flow matching instead of single-shot prediction (ablation) | PuzzleFlow · GAP-3 | Perfect Accuracy (%): 22.6 -> 28.5 | Improved |
| fine-tune the pretrained ViT instead of freezing it (ablation) | PuzzleFlow · GAP-3 | Perfect Accuracy (%): 7.4 -> 28.5 | Improved |
| add 4 transformer layers on top of ViT (vs 0 layers) (ablation) | PuzzleFlow · GAP-3 | Perfect Accuracy (%): 10.1 -> 28.5 | Improved |
| use 4 layers instead of 2 (ablation) | PuzzleFlow · GAP-3 | Perfect Accuracy (%): 23.5 -> 28.5 | Improved |
| use 6 layers instead of 4 (ablation) | PuzzleFlow · GAP-3 | Perfect Accuracy (%): 28.5 -> 24.7 | No clear effect |
| use learned RGBA projection instead of dropping alpha (ablation) | PuzzleFlow · GAP-3 | Perfect Accuracy (%): 9.2 -> 28.5 | Improved |
| add learned RGBA projection (absolute accuracy) (ablation) | PuzzleFlow · GAP-3 | Absolute Accuracy (%): 44.4 -> 62.9 | Improved |
| add flow matching (spatial relationship accuracy) (ablation) | PuzzleFlow · GAP-3 | SRA (%): 50.0 -> 55.7 | Improved |
| PuzzleFlow vs strongest baseline FCViT | PuzzleFlow · GAP-3 | SRA (%): 47.6 -> 55.7 | Improved |
| PuzzleFlow vs DiffAssemble on larger puzzles | PuzzleFlow · GAP-5 | Absolute Accuracy (%): 21.9 -> 29.1 | Improved |
| increase puzzle size from 3x3 to 5x5 | PuzzleFlow · GAP-3 vs GAP-5 | Perfect Accuracy (%): 28.5 -> 0.3 | Hurt |
| apply to square-piece JPwLEG-3 without tuning | PuzzleFlow · JPwLEG-3 | Absolute Accuracy (%): 96.9 -> 72.6 | Hurt |
Closest papers by meaning
- Re-assembling the past: The RePAIR dataset and benchmark for real world 2D and 3D puzzle solving (NeurIPS 2024 · similarity 0.90)
- 3DPGS: 3D Probabilistic Graph Search for Archaeological Piece Grouping (AAAI 2025 · similarity 0.88)
- Solving Masked Jigsaw Puzzles with Diffusion Vision Transformers (CVPR 2024 · similarity 0.86)
- PuzzleFusion++: Auto-agglomerative 3D Fracture Assembly by Denoise and Verify (ICLR 2025 · similarity 0.85)
- Scalable Geometric Fracture Assembly via Co-creation Space among Assemblers (AAAI 2024 · similarity 0.85)
- Puzzles: Unbounded Video-Depth Augmentation for Scalable End-to-End 3D Reconstruction (NeurIPS 2025 · similarity 0.84)
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.
