Research map / Detection and segmentation / Open-set and few-shot detection
ChainHOI: Joint-based Kinematic Chain Modeling for Human-Object Interaction Generation
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 ChainHOI, a diffusion-based text-driven human-object interaction generator that models interactions explicitly at the joint level (HOI joint graph with GST-GCN) and kinetic-chain level (KIM), evaluated on BEHAVE and OMOMO.
Problem
Text-driven human-object interaction generation typically embeds full-body poses as tokens, modeling joint interactions only implicitly, which makes learning hard and yields physically implausible interactions such as penetration and contact errors.
Approach
Builds an HOI joint graph with an object node connected only to potential interaction joints, processes it with a spatiotemporal GCN plus semantic-consistent module, and adds kinetic-chain tokens with masked cross-attention (KIM) so each chain attends only to its own joints, supervised by contact and object-pose losses.
Key results
- On BEHAVE, ChainHOI reaches FID 0.095 versus 0.170 for the next best InterDiff, and R-Top1 0.435 versus 0.457 for HOI-Diff (lower FID is better; HOI-Diff's R-Top1 is higher).
- On OMOMO, ChainHOI achieves FID 0.112 and OCD 0.089 (best), with average inference time 0.61s on BEHAVE versus 3.61s to 38.4s for baselines.
- Ablation on BEHAVE shows removing the kinetic chain attention mask in KIM worsens FID from 0.095 to 0.184.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add joint- and kinetic-chain-level modeling with ChainHOI | text-driven HOI generation · BEHAVE test | FID: 0.157±.001 -> 0.095±.001 | Helped |
| add joint- and kinetic-chain-level modeling with ChainHOI | text-driven HOI generation · OMOMO test | R-Precision Top1: 0.254±.002 -> 0.264±.005 | Helped |
| add joint- and kinetic-chain-level modeling with ChainHOI | text-driven HOI generation · OMOMO test | PS: 0.015±.001 -> 0.019±.001 | Hurt |
| replace GST-GCN and KIM with Transformer (ablation) | ChainHOI · BEHAVE | FID: 0.095 -> 0.229 | Hurt |
| use joint-level representation with Transformer (ablation) | ChainHOI · BEHAVE | FID: 0.095 -> 0.202 | Hurt |
| remove Semantic-consistent Module and KIM (ablation) | ChainHOI · BEHAVE | FID: 0.095 -> 0.193 | Hurt |
| add KIM (ablation) | ChainHOI · BEHAVE | FID: 0.142 -> 0.095 | Helped |
| add Semantic-consistent Module (ablation) | ChainHOI · BEHAVE | FID: 0.170 -> 0.095 | Helped |
| add attention mask in KIM (ablation) | ChainHOI · BEHAVE | FID: 0.184 -> 0.095 | Helped |
| add object geometry (ablation) | ChainHOI · BEHAVE | OCD: 0.278 -> 0.091 | Helped |
Surprising findings
- Removing object geometry improves R-Top1 (0.480) but severely degrades OCD (0.278), suggesting geometry mainly aids fine-grained contact.
Limitations
- Evaluation on only two datasets (BEHAVE and OMOMO) with a newly trained R-Precision feature extractor that is not validated against standard models.
- Contact loss uses only eight selected interaction joints rather than full-mesh SDF, a simplification of penetration modeling.
- Reported OCD relies on ChatGPT 4o to identify semantically consistent sequences.
Experimental setup
Datasets: BEHAVE OMOMO
Benchmarks: BEHAVE OMOMO
Metrics: FID R-Precision OCD PS FSR AIT
Baselines: MDM PriorMDM InterDiff CHOIS HOI-Diff HOI-Diff + AIC
Builds on: ST-GCN HOI-Diff CHOIS MDM DETR
Closest papers by meaning
- CG-HOI: Contact-Guided 3D Human-Object Interaction Generation (CVPR 2024 · similarity 0.90)
- InterAct: Advancing Large-Scale Versatile 3D Human-Object Interaction Generation (CVPR 2025 · similarity 0.89)
- InterDreamer: Zero-Shot Text to 3D Dynamic Human-Object Interaction (NeurIPS 2024 · similarity 0.89)
- Learning to Generate Human-Human-Object Interactions from Textual Descriptions (NeurIPS 2025 · similarity 0.89)
- InteractAnything: Zero-shot Human Object Interaction Synthesis via LLM Feedback and Object Affordance Parsing (CVPR 2025 · similarity 0.88)
- HOI-Dyn: Learning Interaction Dynamics for Human-Object Motion Diffusion (NeurIPS 2025 · 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.