Implicit Hand Reconstruction

Aleksei Zhuravlev, Dr. Danda Pani Paudel, Dr. Thomas Probst

A NeRF-based 3D reconstruction of a human hand from monocular and multi-view sequences, based on Interhand2.6m dataset



Abstract

This work addresses the problem of reconstructing an animatable avatar of a human hand from a collection of images of a user performing a sequence of gestures. Our model can capture accurate hand shape and appearance and generalize to various hand subjects. For a 3D point, we can apply two types of warping: zero-pose canonical space and UV space. The warped coordinates are then passed to a NeRF which outputs the expected color and density. We demonstrate that our model can accurately reconstruct a dynamic hand from monocular or multi-view sequences, achieving high visual quality on the Interhand2.6m dataset.

Method

Architecture of HumanNeRF, adapted to human hand setting instead of full body
Architecture of LiveHand, reimplemented from scratch

Results



Single view multi-pose sequence

Reconstructed avatar in multiple poses from different views

References

  1. LiveHand: Real-time and Photorealistic Neural Hand Rendering. arXiv preprint arXiv:2302.07672
  2. Neuman: Neural human radiance field from a single video. In European Conference on Computer Vision, pp. 402-418. Cham: Springer Nature Switzerland, 2022