Research map / 3D vision / 3D generation
Primitive-Based 3D Human-Object Interaction Modelling and Programming
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.
Introduces P3HAOI, a superquadric-primitive representation and dataset of 23,494 images for single-view 3D reconstruction of humans interacting with articulated objects, with a baseline using an HOI loss for joint optimization.
Problem
3D human-articulated object interaction (HAOI) reconstruction from single images is hard due to scarce 3D ground truth and the heterogeneous representations (SMPL and CAD models) used for humans and objects, which limit flexibility and generalization.
Approach
Represent both humans and articulated objects as compositions of superquadric primitives so that a single shared representation enables mutual information learning across categories, and reconstruct them jointly from a single image with an HOI vector-cosine loss and interpenetration loss.
Key results
- Proposed method reaches mean Chamfer distance of 0.1241 cm for human and 21.4340 cm for object, versus D3D-HOI (2.0135 / 55.3536) and CHORE (33.7820 / 81.8080).
- Ablation shows removing the HOI loss ('no HOI') raises object error on several splits, and the IUV supervision improves part-state prediction.
- Dataset has 23,494 images across 18 articulated object categories, about 1,305 instances per category on average.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use the proposed primitive-based reconstruction method instead of CHORE | Single-view 3D HAOI reconstruction · P3HAOI | Human Chamfer distance (cm): 33.7820 -> 0.1241 | Helped |
| Use the proposed primitive-based reconstruction method instead of CHORE | Single-view 3D HAOI reconstruction · P3HAOI | Object Chamfer distance (cm): 81.8080 -> 21.4340 | Helped |
| Use the proposed primitive-based reconstruction method instead of D3D-HOI | Single-view 3D HAOI reconstruction · P3HAOI | Human Chamfer distance (cm): 2.0135 -> 0.1241 | Helped |
| Use the proposed primitive-based reconstruction method instead of D3D-HOI | Single-view 3D HAOI reconstruction · P3HAOI | Object Chamfer distance (cm): 55.3536 -> 21.4340 | Helped |
| Add IUV supervision to fine-tune image features (ablation) | Ours · P3HAOI | Object Chamfer distance | Helped |
| Add HOI loss for joint human-object optimization (ablation) | Ours · P3HAOI | Object Chamfer distance | Helped |
| Add object 2D keypoint loss (ablation) | Ours · P3HAOI | Reconstruction accuracy | Hurt |
Surprising findings
- Adding object 2D keypoint supervision acts as a distraction and degrades reconstruction, per the authors' ablation.
- CHORE and D3D-HOI, modified as baselines, perform poorly on articulated objects compared with the proposed method.
Limitations
- Hard to predict correct face direction of the object relative to the human in some cases.
- Only 18 articulated object categories and relatively simple objects; more complex articulated objects not covered.
- Baselines are modified adaptations (CHORE, D3D-HOI) rather than native single-view methods.
Experimental setup
Datasets: P3HAOI D3D-HOI SMPL THuman2.0 Mixamo AKB-48 SAPIEN
Benchmarks: P3HAOI
Metrics: Chamfer distance (cm) for human Chamfer distance (cm) for object
Baselines: CHORE D3D-HOI
Builds on: D3D-HOI CHORE ROMP DINOv2 EFT
Closest papers by meaning
- Reconstructing In-the-Wild Open-Vocabulary Human-Object Interactions (CVPR 2025 · similarity 0.86)
- HOI-M^3: Capture Multiple Humans and Objects Interaction within Contextual Environment (CVPR 2024 · similarity 0.86)
- InteractAnything: Zero-shot Human Object Interaction Synthesis via LLM Feedback and Object Affordance Parsing (CVPR 2025 · similarity 0.85)
- InterAct: Advancing Large-Scale Versatile 3D Human-Object Interaction Generation (CVPR 2025 · similarity 0.85)
- End-to-End HOI Reconstruction Transformer with Graph-based Encoding (CVPR 2025 · similarity 0.85)
- Real2Sim in HOI: Toward Physically Plausible HOI Reconstruction from Monocular Videos (NeurIPS 2026 · similarity 0.85)
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.