Research map / Architectures and efficiency / Attention and transformers
PHFormer: Multi-Fragment Assembly Using Proxy-Level Hybrid Transformer
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes PHFormer, a proxy-level hybrid Transformer with intra/inter-fragment attention and an adjacency-aware hierarchical pose estimator for multi-fragment 3D assembly from point clouds.
Problem
Multi-fragment assembly: predicting 6-DoF poses to restore broken objects from geometry alone, without semantic part decomposition, where existing part-assembly methods degrade badly.
Approach
Partition each fragment into proxies (FPS + ball query + EdgeConv), reason over them with alternating intra-fragment and inter-fragment attention, and predict pairwise adjacency and relative poses before integrating them via adjacency-weighted messages into absolute poses.
Key results
- On Breaking Bad Everyday (at most 20 fragments), rotation MAE of 22.4 degrees vs. 41.1 for Jigsaw and 66.5-72.4 for other baselines, with 10.7 ms inference vs. 3048.4 ms for Jigsaw.
- Part accuracy 50.7% vs. 51.6% for Jigsaw (Jigsaw slightly higher on PA).
- Generalization on the Other subset: rotation MAE 39.1 degrees when trained on Everyday, vs. 66.5+ for baselines in Table 3 comparisons.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use PHFormer instead of Jigsaw | fragment assembly · Breaking Bad Everyday, ≤20 fragments | Rotation MAE (degree): 41.1 -> 22.4 | Helped |
| Use PHFormer instead of Jigsaw | fragment assembly · Breaking Bad Everyday, ≤20 fragments | Inference time (ms): 3048.4 -> 10.7 | Helped |
| Fine-tune on Everyday before training on Artifact | PHFormer · Breaking Bad Artifact | Rotation MAE (degree): 31.6 -> 27.7 | Helped |
| Replace PointNet with EdgeConv (ablation) | baseline assembly model · Breaking Bad Everyday, ≤20 fragments | Rotation MAE (degree): 44.2 -> 36.9 | Helped |
| Replace absolute pose estimator with hierarchical pose estimator (ablation) | PointNet and intra-attention baseline · Breaking Bad Everyday, ≤20 fragments | Rotation MAE (degree): 44.2 -> 41.6 | Helped |
| Use proxy-level features and fragment-aware attention instead of fragment-level features and intra-attention (ablation) | EdgeConv model with hierarchical pose estimator · Breaking Bad Everyday, ≤20 fragments | Rotation MAE (degree): 36.9 -> 24.1 | Helped |
| Replace fragment-aware attention with hybrid attention (ablation) | proxy-feature model with hierarchical pose estimator · Breaking Bad Everyday, ≤20 fragments | Rotation MAE (degree): 24.1 -> 24.4 | Hurt |
| Use a hierarchical pose estimator instead of an absolute pose estimator (ablation) | EdgeConv proxy model with hybrid attention · Breaking Bad Everyday, ≤20 fragments | Rotation MAE (degree): 24.4 -> 23.0 | Helped |
| Add adjacency prediction to the hierarchical pose estimator (ablation) | EdgeConv proxy model with hybrid attention · Breaking Bad Everyday, ≤20 fragments | Rotation MAE (degree): 23.0 -> 22.4 | Helped |
Surprising findings
- Jigsaw achieves higher part accuracy (51.6%) than PHFormer (50.7%) despite worse rotation and translation errors.
- The semantic part assembly baseline RGL-Net performs poorly even when given a sorted fragment order prior.
Limitations
- Performance degrades or the challenge remains for objects with more than 20 fragments, as stated by the authors.
- Relies on a large-scale synthetic benchmark (Breaking Bad) with simulated fractures; real-world archaeological fragments not evaluated.
Experimental setup
Datasets: Breaking Bad
Benchmarks: Breaking Bad
Metrics: rotation RMSE rotation MAE translation RMSE translation MAE Chamfer distance part accuracy inference time
Baselines: Global LSTM DGL RGL-Net SGSA Jigsaw
Builds on: Transformer EdgeConv PointNet++ Breaking Bad benchmark
Closest papers by meaning
- Generative 3D Part Assembly via Part-Whole-Hierarchy Message Passing (CVPR 2024 · similarity 0.86)
- Scalable Geometric Fracture Assembly via Co-creation Space among Assemblers (AAAI 2024 · similarity 0.85)
- Neural Assembler: Learning to Generate Fine-Grained Robotic Assembly Instructions from Multi-View Images (AAAI 2025 · similarity 0.85)
- Imagine: Image-Guided 3D Part Assembly with Structure Knowledge Graph (AAAI 2025 · similarity 0.84)
- Twinner: Shining Light on Digital Twins in a Few Snaps (CVPR 2025 · similarity 0.84)
- DiffAssemble: A Unified Graph-Diffusion Model for 2D and 3D Reassembly (CVPR 2024 · similarity 0.83)
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.