Research map / 3D vision / Face and shape
BiTT: Bi-directional Texture Reconstruction of Interacting Two Hands from a Single Image
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 BiTT, a per-scene, end-to-end trainable method that reconstructs relightable, pose-free UV textures of two interacting hands from a single RGB image using left/right texture symmetry and the HTML texture parametric model.
Problem
Reconstructing personalized, relightable hand textures (including invisible sides) of two interacting hands from only a single RGB image, where prior methods need multi-view images or videos and handle only single hands.
Approach
A bi-directional texture reconstructor exploits left/right hand texture symmetry (cross-hand skip features) and a PCA-based hand texture parametric model (HTML) supplies a coarse full texture for occluded pixels, followed by Phong-rendered fine-stage refinement with albedo-consistency and symmetry losses.
Key results
- On InterHand2.6M with GT mesh, BiTT reaches PSNR 30.41 and LPIPS 0.1019 on appearance reconstruction, versus HARP's 28.11 and 0.0696 (LPIPS favors HARP).
- On novel views with GT mesh, BiTT achieves PSNR 27.79 vs HARP 25.97 and S2Hand 25.73.
- Without GT mesh on InterHand2.6M, BiTT gets PSNR 28.40 vs HARP 25.17 on appearance reconstruction.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use BiTT instead of HARP | hand texture reconstruction · InterHand2.6M, appearance reconstruction (GT mesh) | PSNR: 28.11 -> 30.41 | Helped |
| use BiTT instead of HARP | hand texture reconstruction · InterHand2.6M, novel poses (GT mesh) | PSNR: 25.79 -> 26.54 | Helped |
| use BiTT instead of HARP | hand texture reconstruction · InterHand2.6M, different views (GT mesh) | PSNR: 25.97 -> 27.79 | Helped |
| use BiTT instead of S2Hand | hand texture reconstruction · InterHand2.6M, appearance reconstruction (no GT mesh) | PSNR: 25.72 -> 28.40 | Helped |
| use BiTT instead of HARP | hand texture reconstruction · InterHand2.6M, novel poses (no GT mesh) | PSNR: 24.32 -> 25.12 | Helped |
| use BiTT instead of HARP | hand texture reconstruction · InterHand2.6M, different views (no GT mesh) | PSNR: 24.49 -> 26.34 | Helped |
| use BiTT instead of S2Hand | hand texture reconstruction · RGB2Hands, appearance reconstruction | PSNR: 25.72 -> 26.02 | Helped |
| use BiTT instead of S2Hand | hand texture reconstruction · RGB2Hands, novel poses | PSNR: 24.22 -> 24.54 | Helped |
| add symmetric texture reconstruction (ablation) | BiTT without symmetric texture reconstruction and albedo consistency loss · InterHand2.6M, novel poses and viewpoints | PSNR: 25.36 -> 26.21 | Helped |
| add albedo consistency loss (ablation) | BiTT without albedo consistency loss · InterHand2.6M, novel poses and viewpoints | PSNR: 26.21 -> 27.16 | Helped |
| add albedo consistency loss (ablation) | BiTT without albedo consistency loss · InterHand2.6M, novel poses and viewpoints | LPIPS: 0.1230 -> 0.1176 | Helped |
Surprising findings
- HARP achieves better LPIPS than BiTT on several InterHand2.6M and RGB2Hands settings despite lower PSNR, and BiTT's gain on RGB2Hands is smaller because of lower usable symmetric texture.
Limitations
- Seriously misaligned meshes still cause significant noise in reconstructed textures.
- Tattoos or accessories on invisible hand sides are not handled.
- Requires per-scene training rather than generalizing to new inputs.
Experimental setup
Datasets: InterHand2.6M RGB2Hands
Benchmarks: InterHand2.6M RGB2Hands
Metrics: L1 LPIPS PSNR SSIM MS-SSIM
Baselines: S2Hand HTML HARP HandNeRF AMVUR
Builds on: HTML NeRF MANO Phong illumination
Closest papers by meaning
- Fine-Grained Multi-View Hand Reconstruction Using Inverse Rendering (AAAI 2024 · similarity 0.88)
- BIGS: Bimanual Category-agnostic Interaction Reconstruction from Monocular Videos via 3D Gaussian Splatting (CVPR 2025 · similarity 0.88)
- HOLD: Category-agnostic 3D Reconstruction of Interacting Hands and Objects from Video (CVPR 2024 · similarity 0.87)
- SRHand: Super-Resolving Hand Images and 3D Shapes via View/Pose-aware Neural Image Representations and Explicit Meshes (NeurIPS 2025 · similarity 0.87)
- EasyHOI: Unleashing the Power of Large Models for Reconstructing Hand-Object Interactions in the Wild (CVPR 2025 · similarity 0.86)
- HOGSA: Bimanual Hand-Object Interaction Understanding with 3D Gaussian Splatting Based Data Augmentation (AAAI 2025 · 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.