Annonce

Les commentaires sont clos.

Deformable 3D Reconstruction Ambiguities

26 Juillet 2022


Catégorie : Post-doctorant


The Encov team at Institut Pascal, Clermont-Ferrand, France is looking for a postdoc student to work on deformable 3D reconstruction ambiguities (Link to offer). The goal of this project is to uncover the disambiguates in deformable 3D reconstruction in order to self-calibrate the camera.

The postdoc will be supervised by Prof. Adrien Bartoli (adrien.bartoli@gmail.com) and Shaifali Parashar (shaifali.parashar@liris.cnrs.fr).

Requirements:

  1. Strong background in computer vision and mathematics with a strong publishing record
  2. Strong programming skills in C++ and python
  3. Fluency in English

Project duration: 12 months (with a possible extension)

Tentative start date: September 2022

How to apply:

Please send your CV and list of publications to Prof. Adrien Bartoli and Shaifali Parashar with the subject "Postdoc position in Deformable 3D Reconstruction Ambiguities".

 

A detailed understanding of the rigid 3D scenes through images can be formed easily, thanks to the development of Multiple View Geometry (MVG[1]) in the 90s. The case for deformable objects is still an open research problem. It involves an additional modelling of deformations which makes it rather complicated. Numerous strategies have been developed in the past 30 years, ranging from statistical approximations to physical ones (constraining lengths and other metric quantities), all missing a concrete, mathematical way of formulating deformations. In our recent works[2], we have shown that breaking a deformable scene into infinitesimal planes allows the deformations to be considered as linear functions which leads to simple, fast solutions to their reconstruction through images. This brought a substantial improvement to the state of the art in terms of complexity and accuracy. With such modelling, we have also shown that it is possible to self-calibrate[3] (solution limited to estimation of focal length only) through a deformable scene. In the case of rigid objects, it is known that there are certain camera motions, also known as critical motion sequences, that prevent self-calibration as they cause an ambiguous computation of calibration parameters. However, the ambiguities in the case of self-calibration of deformable objects are yet to be uncovered.

In this project we aim to take a step towards developing a concrete understanding of deformable scenes by uncovering the camera and structure ambiguities. Given that a deformable object is infinitesimally planar, our goal is to study the possible geometric configurations of an object that image alike through the uncalibrated cameras. To be specific, there is a range of homographic transformations, represented by the absolute quadric, that can be applied to an object while keeping the image as a constant. By restricting the deformation of the object, we can impose constraints to disambiguate the possible configurations and uncover the ambiguous cases. Our goal is to explore these ambiguities to recover cameras (relatively) in case of deformable scenes. Contributions are expected to the differential modelling of geometric vision and concrete methods to advanced Shape-from-Template and Non-Rigid Structure-from-Motion. This would be beneficial in improving Visual SLAM in deformable scenes[4].

 

[1] Hartley and Zissermann. Multiple View Geometry.

[2] Parashar et al, TPAMI 2017. Isometric Non-Rigid Shape-from-Motion with Riemannian Geometry in Linear Time.

[3] Parashar et al, ECCV, 2018. Self-Calibrating Isometric Non-Rigid Structure-from-Motion.

[4] Lamarca et al, TRO 2020. DefSLAM: Tracking and Mapping of Deforming Scenes from Monocular Sequences.