Enhancing Monocular 3D Hand Reconstruction with Learned Texture Priors

Official repository for the paper:

Enhancing Monocular 3D Hand Reconstruction with Learned Texture Priors
Giorgos Karvounas, Nikolaos Kyriazis, Iason Oikonomidis, Georgios Pavlakos, Antonis A. Argyros
arXiv:2508.09629

---

📌 Overview

This work revisits the role of texture in monocular 3D hand reconstruction, treating it not only as a tool for photorealism but as a dense, spatially grounded supervisory signal that enhances pose and shape estimation.

We propose a lightweight, transformer-based texture module that consolidates sparse UV–RGB observations into a full texture prior. Integrated into existing pipelines (e.g., HaMeR), our approach delivers measurable accuracy and realism gains, particularly in occluded and egocentric scenarios, without adding any test-time overhead.

---

🚀 Key Contributions

• Introduces the first unified framework for learning texture priors from sparse, monocular observations.
• Transformer-based module with pixel-level attention for coherent texture reconstruction.
• Weakly-supervised training using differentiable rendering — no manual annotations required.
• Improves state-of-the-art monocular hand reconstruction benchmarks (e.g., +2.7% PCK on occluded hands).

---

🔧 Setup Pretrained Weights

1. Download the weights
   👉 texture_supervised_hamer_weights These weights have been retrained with our texture–supervised framework, incorporating the learned texture priors described in the paper.

2. Place the file into the following directory (after cloning and setting up the HaMeR repo): hamer/_DATA/hamer_ckpts/checkpoints/

3. **Update the init file with hamer.ckpt -> texture_supervised_hamer_weights.ckpt

---

📢 Updates

• Release code
• Add preparation instructions
• Provide demo notebooks
• Release new weights for HaMeR