Research map / 3D vision / 3D generation
HaNDF: Object-Conditioned Geometric Neural Hand Distance Fields
Zuhao Liu, Melih Darcan, Zhengdi Yu, Riza Alp Guler, Tolga Birdal
Read the paperCheck your idea against 137,217 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 HaNDF, object-conditioned neural distance fields serving as data-driven priors over plausible hand-object configurations in the product manifold of hand articulations.
Problem
Unconditional generative priors model human pose well, but hands interact with objects, so a generative prior conditioned on objects is needed to model plausible hand-object configurations.
Approach
Object-conditional neural distance fields defined over the product manifold of hand articulations, used as data-driven priors for hand-object configurations, extending unconditional pose priors to object-conditioned ones.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use Riemannian distance field instead of Euclidean (ablation) | HaNDF (MLP + object condition) · HOI4D denoising sigma=0.8 | delta_q (degrees): 8.15 -> 5.98 | Improved |
| use Riemannian field instead of Euclidean with transformer and object condition (ablation) | HaNDF (Transformer) · HOI4D denoising sigma=0.8 | delta_q (degrees): 10.27 -> 5.52 | Improved |
| add object condition (object token) (ablation) | HaNDF (Transformer, Riemannian) · HOI4D denoising sigma=0.8 | delta_q (degrees): 6.10 -> 5.52 | Improved |
| condition on the correct object instead of a random wrong-category object (ablation) | HaNDF · HOI4D denoising sigma=0.8 | delta_q (degrees): 5.80 -> 5.52 | Improved |
| use transformer instead of MLP backbone with object condition and Riemannian field (ablation) | HaNDF · HOI4D denoising sigma=0.8 | delta_q (degrees): 5.98 -> 5.52 | Improved |
| use DeepSDF-style auto-decoder object code instead of COD-VAE or Tencent-XR-3DGen (ablation) | HaNDF · HOI4D denoising sigma=0.8 | delta_q (degrees): 6.11 -> 5.52 (COD-VAE); 5.70 -> 5.52 (3DGen) | Improved |
| encode object code as 1 token instead of 64 tokens (ablation) | HaNDF · HOI4D denoising sigma=0.8 | delta_q (degrees): 6.40 -> 5.52 | Improved |
| increase transformer depth from 4 to 8 layers (ablation) | HaNDF · HOI4D denoising sigma=0.8 | dNN (degrees): 4.96 -> 5.07 | Hurt |
| increase transformer depth from 2 to 4 layers (ablation) | HaNDF · HOI4D denoising sigma=0.8 | delta_q and dNN (degrees): delta_q 4.95 -> 5.52; dNN 6.08 -> 4.96 | Mixed |
| plug HaNDF into AlignSDF as hand refinement prior | AlignSDF · HOI4D hand-object recovery | delta_q (degrees): 14.77 -> 9.60 | Improved |
| add nearest-neighbor object retrieval in DeepSDF latent space | AlignSDF + HaNDF · HOI4D hand-object recovery | obj-F@2%: 0.1187 -> 0.1568 | Improved |
| use K=8 object optimization instead of K=1 retrieval | AlignSDF + HaNDF · HOI4D hand-object recovery | obj-CD: obj-CD 0.0896 -> 0.1083; MPJPE 11.50 -> 11.45 | Mixed |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- G-HOP — G-HOP: Generative Hand-Object Prior for Interaction Reconstruction and Grasp Synthesis (CVPR 2024) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- G-HOP: Generative Hand-Object Prior for Interaction Reconstruction and Grasp Synthesis (CVPR 2024 · similarity 0.89)
- NRDF: Neural Riemannian Distance Fields for Learning Articulated Pose Priors (CVPR 2024 · similarity 0.89)
- HOISDF: Constraining 3D Hand-Object Pose Estimation with Global Signed Distance Fields (CVPR 2024 · similarity 0.88)
- GEARS: Local Geometry-aware Hand-object Interaction Synthesis (CVPR 2024 · similarity 0.88)
- HandBooster: Boosting 3D Hand-Mesh Reconstruction by Conditional Synthesis and Sampling of Hand-Object Interactions (CVPR 2024 · similarity 0.87)
- MEgoHand: Multimodal Egocentric Hand-Object Interaction Motion Generation (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.
