Research map / Image generation / Image restoration
Image Reassembly Combining Deep Learning and Shortest Path Problem
Marie-Morgane Paumard, David Picard, Hedi Tabia
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 |
|---|---|---|---|
| fuse fragment features with Kronecker product instead of concatenation (ablation) | relative position CNN · ImageNet val (8-class position) | accuracy (%): 64.6% -> 66.4% | Improved |
| fuse with Kronecker product instead of Hadamard low-rank approximation (ablation) | relative position CNN · ImageNet val (8-class position) | accuracy (%): 59.2% -> 66.4% | Improved |
| use bilinear/Kronecker fusion network instead of Doersch et al. reproduction | relative position CNN · ImageNet val (8-class position) | accuracy (%): 57.0% -> 66.4% | Improved |
| solve the 9-class joint problem instead of the sequential 2-class then 8-class approach | position classifier · ImageNet val | accuracy (%): 66.4% -> 64.2% | No clear effect |
| use Dijkstra shortest path instead of greedy assignment (ablation) | reassembly, central known, complete puzzle · MET reassembly | reconstruction accuracy (%): 41.0 -> 44.4 | Improved |
| use Dijkstra instead of greedy when central fragment unknown (ablation) | reassembly, central unknown · MET reassembly | reconstruction accuracy (%): 36.2 -> 39.2 | Improved |
| use Dijkstra instead of greedy with incomplete puzzle (ablation) | reassembly, incomplete · MET reassembly | reconstruction accuracy (%): 26.5 -> 29.5 | Improved |
| use Dijkstra instead of greedy for position accuracy, central known (ablation) | reassembly · MET reassembly | position accuracy (%): 87.7 -> 89.9 | Improved |
| solve global assignment instead of per-fragment classifier output | reassembly · MET reassembly | position accuracy (%): 66.4% -> around 90% | Improved |
| make the central fragment unknown | Dijkstra reassembly · MET reassembly | reconstruction accuracy (%): 44.4 -> 39.2 | Hurt |
| add one outsider fragment to the puzzle | Dijkstra reassembly · MET reassembly | reassembly accuracy (%): 44.4% -> 26.3% | Hurt |
| add two outsider fragments to the puzzle | Dijkstra reassembly · MET reassembly | reassembly accuracy (%): 44.4% -> 14.3% | Hurt |
Closest papers by meaning
- Deep Model Reassembly (NeurIPS 2022 · similarity 0.88)
- DiffAssemble: A Unified Graph-Diffusion Model for 2D and 3D Reassembly (CVPR 2024 · similarity 0.88)
- Beyond Reassembly: Fractured Object Recovery with Missing Parts (CVPR 2026 · similarity 0.88)
- ReassembleNet: Learnable Keypoints and Diffusion for 2D Fresco Reconstruction (ICCV 2025 · similarity 0.87)
- PHFormer: Multi-Fragment Assembly Using Proxy-Level Hybrid Transformer (AAAI 2024 · similarity 0.84)
- Model LEGO: Creating Models Like Disassembling and Assembling Building Blocks (NeurIPS 2024 · 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.
