Research map / 3D vision / Face and shape
ContactField: Implicit Field Representation for Multi-Person Interaction Geometry
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 ContactField, an implicit field model that jointly predicts occupancy, instance ID, and contact fields for close multi-person interactions from multi-view images, and releases a synthetic SynMPI dataset of 2-4 person interactions.
Problem
Reconstructing geometry, instance identity, and contact of closely interacting humans from multi-view images is hard due to occlusion, inaccurate geometry, and scarce multi-person training data.
Approach
An implicit occupancy-plus-ID field built on a multi-view local-global feature module (view-to-view transformer for local features, SRT-style set-latent scene encoder and decoder for global features), with contact estimated unsupervised as the local standard deviation of predicted ID values.
Key results
- On Hi4D, ContactField reaches CD 0.406, P2S 0.329, NC 0.892 versus DeepMultiCap's CD 0.631, P2S 0.495, NC 0.768.
- Removing the SRT global feature raises Hi4D CD to 0.468 and P2S to 0.402, showing the global features help.
- Variance-based contact estimation achieves CP 0.629 at epsilon 0.05, versus 0.518 for mesh-based contact inference.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add SRT global scene features (ablation) | ContactField · Hi4D test | CD: 0.468 -> 0.406 | Helped |
| add SRT global scene features (ablation) | ContactField · Hi4D test | contact precision (ϵ=0.075): 0.458 -> 0.670 | Helped |
| omit both grouping-loss terms (ablation) | ContactField · Hi4D test | contact precision (ϵ=0.075): 0.670 -> 0.187 | Hurt |
| use squared-distance grouping term only (ablation) | ContactField · Hi4D test | CD: 0.532 -> 0.400 | Helped |
| use exponential-penalty grouping term only (ablation) | ContactField · Hi4D test | CD: 0.400 -> 0.532 | Hurt |
| add the exponential-penalty term alongside squared distance (ablation) | ContactField · Hi4D test | contact precision (ϵ=0.075): 0.528 -> 0.670 | Helped |
| increase input views from 4 to 8 (ablation) | ContactField · Hi4D test | CD: 0.761 -> 0.406 | Helped |
| estimate contact using ID-field variance instead of output meshes (ablation) | ContactField · Hi4D test | contact precision (ϵ=0.075): 0.621 -> 0.670 | Helped |
| add SynMPI training data (ablation) | ContactField · Hi4D test | CD: 0.499 -> 0.406 | Helped |
| use LVD instead of MVPose for SMPL initialization (ablation) | DMC · SynMPI test | CD: 0.805 -> 0.631 | Helped |
Surprising findings
- Adding the synthetic SynMPI data improves Hi4D test results (CD 0.499 to 0.406), indicating benefit from synthetic training data.
Limitations
- Without a spatial prior such as SMPL, ID prediction fails in extreme poses such as hugging.
- Resolution constraints when reconstructing geometry, ID and contact jointly limit fine detail.
- Contact precision depends strongly on 3D grid resolution.
Experimental setup
Datasets: Hi4D SynMPI
Benchmarks: Hi4D SynMPI
Metrics: Chamfer Distance Point-to-Surface Normal Consistency Contact Precision
Baselines: DeepMultiCap
Builds on: PIFu DoubleField Scene Representation Transformer DeepMultiCap associative embedding
Closest papers by meaning
- InteractVLM: 3D Interaction Reasoning from 2D Foundational Models (CVPR 2025 · similarity 0.90)
- Pose Priors from Language Models (CVPR 2025 · similarity 0.88)
- InterAct: Advancing Large-Scale Versatile 3D Human-Object Interaction Generation (CVPR 2025 · similarity 0.87)
- PICO: Reconstructing 3D People In Contact with Objects (CVPR 2025 · similarity 0.87)
- Capturing Closely Interacted Two-Person Motions with Reaction Priors (CVPR 2024 · similarity 0.87)
- Template Free Reconstruction of Human-object Interaction with Procedural Interaction Generation (CVPR 2024 · similarity 0.87)
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.