Internship - Characterizing and Optimizing Multi-View Vision Systems

Context

Context

Multi-view reconstruction aims to recover the 3D geometry of a scene from images captured by multiple cameras (e.g.,
[1]
). By combining observations of the same scene points from different viewpoints, multi-view methods can estimate depth and reconstruct detailed 3D surfaces. Reconstruction quality strongly depends on the camera configuration, including camera positions, orientations, focal lengths, fields of view, and image resolutions. Together, these parameters determine which parts of the scene are visible, how accurately they are observed, and how effectively their 3D geometry can be recovered. However, a general framework for quantitatively characterizing the ability of a multi-camera system to reconstruct a given 3D scene is still lacking.
Recent work on camera placement and next-best-view selection has primarily focused on
coverage and observation quality
. In particular,
[2]
investigates camera placement using learned observation fields, while
[3]
proposes selecting informative views by prioritizing geometry that is insufficiently covered by existing cameras. Although these approaches provide useful tools for optimizing camera configurations, they do not establish a general framework for quantitatively assessing the reconstruction capabilities of a multi-view vision system.
[1]
ProbeSDF
Toussaint et al., CVPR 2025
[2]
Neural Observation Field Guided Hybrid Optimization of Camera Placement
Cao et al., IEEE Robotics and Automation Letters 2024
[3]
Coverage Optimization for Camera View Selection
Chen et al., CVPR 2026
[4]
Optimal Camera Placement for Motion Capture Systems
Rahimian & Kearney, IEEE TVCG 2017
Team and Supervision
The internship will be carried out within the
Kinetix
team at INRIA Grenoble Rhône-Alpes. The student will be supervised by
Edmond Boyer
and
Julien Pansiot
.

Assignment

Internship
The goal of this internship is to develop a general framework for characterizing the observation capabilities of multi-view systems.
Metrics that will be investigated include:
the number of pixels observing each voxel, leading to a 3D observation density;
the number of pixels observing each surface element, leading to a surface observation density;
multi-view redundancy and viewing-direction distributions;
potentially, metrics related to geometric reconstruction quality.

Main activities

Part of the internship will be the development of a
reusable library
implementing these metrics together with interactive 2D/3D visualization tools for analyzing camera configurations and identifying poorly observed regions.
A more exploratory objective will be to use these metrics for
camera-system optimization
. Given a target scene and an objective, the system could optimize camera positions, orientations, focal lengths, and resolutions to maximize coverage, observation density, or another task-specific criterion. This connects the project to recent
work [
3]on camera placement and active view selection.
The project combines multi-view geometry, 3D vision, optimization, and software development, with the possibility of pursuing research directions depending on the results.

Skills

Profile

We are looking for a highly motivated Master student M2 in Computer Science, Computer Vision, Applied Mathematics, Robotics, Computer Graphics, or a related field. Good programming skills in Python and/or C++ are required. Background in computer vision, computer graphics, robotics, or 3D geometry would be appreciated.