2020-03154 - Combining implicit and explicit surface representations for 3D human reconstruction from a single image

Contract type : Internship agreement

Level of qualifications required : Graduate degree or equivalent

Fonction : Internship Research

Level of experience : Recently graduated

About the research centre or Inria department

The Inria Rennes - Bretagne Atlantique Centre is one of Inria's eight centres and has more than thirty research teams. The Inria Center is a major and recognized player in the field of digital sciences. It is at the heart of a rich R&D and innovation ecosystem: highly innovative PMEs, large industrial groups, competitiveness clusters, research and higher education players, laboratories of excellence, technological research institute, etc.

Assignment

Keywords: Human 3D reconstruction, parametric shape models, implicit surface representation.
Research fields: Deep learning, 3D computer vision.

3D reconstruction of humans is a popular problem in computer vision and graphics. While earlier successful methods in the field relied on triangulation from multiple cameras or depth sensors for estimating the 3D, learning based approaches have recently allowed to lower the acquisition constraints. Nowadays, many deep learning based methods can recover 3D models of humans from a single color image, by learning strong statistical priors from substantial amounts of training data.

A family of these works use parametric shape models [1,2] and/or explicit surfaces [3,4] to represent the underlying naked human body. They tend to generalize well to images in the wild and they mostly only require weak supervision in training in the form of body joint locations. Another line of work represents the clothed human model with a learnable implicit function [5,6,7]. These latter methods allow to recover detailed surfaces at higher resolutions and can encode varying typologies of clothing. However, they require full 3D ground-truth models for training supervision and do not generalize as well in the wild.

The goal of this internship is to research these surface representations and combine them to allow the best of both worlds for the task of 3D human reconstruction from a single image within a deep learning framework.

References:
[1] Learning to Reconstruct 3D Human Pose and Shape via Model-fitting in the Loop, Kolotouros et al., ICCV 2019
[2] VIBE: Video Inference for Human Body Pose and Shape Estimation, Kocabas et al., CVPR 2020
[3] I2L-MeshNet: Image-to-Lixel Prediction Network for Accurate 3D Human Pose and Mesh Estimation from a Single RGB Image, Moon et al., ECCV 2020
[4] 3D Human Mesh Regression with Dense Correspondence, Zeng et al., CVPR 2020
[5] PIFu: Pixel-Aligned Implicit Function for High-Resolution Clothed Human Digitization, Saito et al., ICCV 2019.
[6] Monocular Real-Time Volumetric Performance Capture, Li et al, ECCV 2020
[7] ARCH: Animatable Reconstruction of Clothed Humans, Huang et la., CVPR 2020

Main activities

  • Participating in the research discussions and algorithm design.
  • Reading and implementing research papers.
  • Reproducing state-of-the-art results.
  • Implementing the ideas proposed by the research collaborators.
  • Creating training and testing datasets.
  • Participating in the publication of the research results.

Skills

Candidates should be preparing a MSc or equivalent degree in computer science, applied mathematics, computer vision, computer graphics or machine learning. Proficiency in coding in Python / Pytorch is a plus.

Benefits package

  • Subsidized meals
  • Partial reimbursement of public transport costs

Remuneration

Remuneration according to the current hourly rate