Back to Feed
Computer Vision / Efficiency & Inference

Creating Realistic 3D Hand Models from Photos

Original: OASIS: Occlusion-aware Single-image Hand Avatar Reconstruction via 3D Gaussian Splatting
By Zhisheng Han, Shiyao Wu, Jiayan Qiu, Yakun Ju, Lu Liu, Le Zhang, Pengfei Feng, Huiyu Zhou, Zheheng Jiang Published 2026-07-31

Listen to the summary

Uses a voice available on your device

Audio options
On this page

Key Takeaways

  • The method solves the difficulty of reconstructing 3D hands from flat images by intelligently managing how parts of the hand hide behind others.
  • It uses a new Feature-on-Mesh approach that tracks how hand surfaces stretch and move, leading to better visual detail than older methods.
  • The system is highly efficient, capable of rendering 3D hand animations at 390 frames per second.
  • It supports a quick one-shot adaptation process that takes about five minutes to customize the model to a specific target image.

Summary & Methodology Analysis

The researchers developed a technique called OASIS to turn a single photograph into a movable 3D hand model. Because a single picture lacks depth, it is hard to guess what the back of a hand or hidden fingers look like. To fix this, the system uses a digital hand skeleton known as a template and projects image details onto 3D points. It uses a special technique called Visibility-conditioned Point-Image Attention, which acts like a smart filter to decide whether a point on the hand is visible or hidden. If a part of the hand is visible, the system prioritizes local visual data, but if it is blocked, it uses global information to make a logical guess about the hidden appearance.

Interactive System Flowchart

Click diagram to expand and zoom

Cross-Examination & FAQs

A deeper dive clarifying mechanics, constraints, and baseline evaluations.

Q1. What is the main goal of this research?

The goal is to create high-quality, 3D animated hand avatars starting from only a single input image.

Q2. Why is it difficult to make 3D hands from 2D photos?

It is difficult because hands have complex shapes that change when they move, and parts of the hand often block others from view, leaving gaps in the information available in a photo.

Q3. How fast does this system run?

The system is capable of rendering at approximately 390 frames per second.

Q4. What is the Feature-on-Mesh representation?

This is a method where the system calculates visual features directly on the flat triangular surfaces that make up the 3D hand model, rather than just on individual points, which helps the model better reflect how skin stretches during movement.

Q5. How does the model handle occlusions or hidden parts of the hand?

It uses a depth test to estimate visibility and then adjusts its attention mechanisms to either look for local visual evidence when a point is clear or shift to global information when the point is hidden.

Q6. What happens during the one-shot adaptation phase?

The system takes about five minutes to perform color calibration and fine-tuning on the specific target image to ensure the final 3D model accurately reflects the texture and details of the person in the photo.

Q7. What datasets were used to train the model?

The model was pretrained using 21 subjects from the InterHand2.6M dataset.

Q8. Does the paper describe the limitations of this method?

The paper includes an appendix regarding limitations and failure cases, though the specific contents of that section are not provided in the extracted text.

Q9. How does this method compare to older techniques like OHTA?

The paper notes that older methods like OHTA are often computationally expensive and struggle to capture fine details because they do not connect the image data and the 3D structure as effectively.