Research map / Detection and segmentation / Open-set and few-shot detection
Implicit Modeling of Non-rigid Objects with Cross-Category Signals
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 MODIF, a multi-object deep implicit function that jointly learns per-category deformation fields, latent codes, and a cross-category refinement field with an attraction-repulsion contact loss for non-interpenetrating organs.
Problem
Deep implicit functions have been limited to single-object scenarios, and multi-object methods ignore interrelations between categories, causing interpenetration of non-rigid objects such as organs.
Approach
Decomposes a multi-object instance into per-category sub-functions (template, hyper-net-conditioned deformation/correction fields) whose features feed a shared cross-category refinement MLP that corrects SDFs, supervised by an attraction-repulsion contact loss and a generalized rigid-plus-non-rigid deformation formulation.
Key results
- On WORD, MODIF achieves mean CD 14.63 and IV 4.24 versus 16.68 CD and 11.39 IV for the best prior object-level baseline (DeepSDF-style row), and 24.28 CD for NDF.
- On MMWHS missing-object recovery (left-myocardium), MODIF reaches mean CD 45.52 on the missing organ versus 94.44 for DIF and 142.19 for NDF.
- Ablation on WORD shows removing the cross-category correction and contact loss raises CD from 14.63 to 15.90, and removing the cross-category correction alone raises EMD from 8.00 to 9.17.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add cross-category refinement and contact loss (ablation) | MODIF · WORD | CD: 15.90 -> 14.63 | Helped |
| add contact loss (ablation) | MODIF · WORD | CD: 16.36 -> 14.63 | Helped |
| add cross-category refinement (ablation) | MODIF · WORD | CD: 15.33 -> 14.63 | Helped |
| add non-rigid deformation Δp_j (ablation) | MODIF · WORD | CD: 15.06 -> 14.63 | Helped |
| add rigid transformation (r_i; t_i) (ablation) | MODIF · WORD | CD: 15.44 -> 14.63 | Helped |
| use MODIF instead of NDF | multi-object reconstruction · MMWHS | CD mean: 4.37 -> 4.95 | Hurt |
| use MODIF instead of DIF | multi-object reconstruction · WORD | IV mean: 16.66 -> 4.24 | Helped |
| use MODIF instead of DIF | missing-object recovery · MMWHS missing left-myocardium | CD mean (missing organ): 94.44 -> 45.52 | Helped |
Surprising findings
- Merging all organs into one instance-level field (DIF, NDF, DeepSDF) performs far worse than object-level modeling, with DIF CD 72.96 on WORD instance-level merged meshes.
Limitations
- Requires class-labeled shapes, so it cannot handle unlabeled or open-set instances.
- The deformation function is overly generic; category-specific priors are not injected.
- Evaluated only on three medical datasets and a small procedural chairs toy set.
Experimental setup
Datasets: WORD AbdomenCT-1K MMWHS
Benchmarks: WORD AbdomenCT-1K MMWHS
Metrics: Chamfer distance earth-mover distance intersection volume
Baselines: DeepSDF DIT DIF NDF DMM (ada)
Builds on: DIF-Net DeepSDF Deep Implicit Templates DMM
Closest papers by meaning
- SPAMs: Structured Implicit Parametric Models (CVPR 2022 · similarity 0.89)
- Deformed Implicit Field: Modeling 3D Shapes With Learned Dense Correspondence (CVPR 2021 · similarity 0.89)
- Deep Implicit Templates for 3D Shape Representation (CVPR 2021 · similarity 0.88)
- Local Deep Implicit Functions for 3D Shape (CVPR 2020 · similarity 0.88)
- $\texttt{NAISR}$: A 3D Neural Additive Model for Interpretable Shape Representation (ICLR 2024 · similarity 0.87)
- Neural Pose Representation Learning for Generating and Transferring Non-Rigid Object Poses (NeurIPS 2024 · similarity 0.86)
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.