Context

Reconstructing urban scenes in 3D with compact mesh-based representations from satellite images has been a long-standing problem in remote sensing and computer vision. Many works have focused their effort on producing 3D models with a LOD1 CityGML formalism [1], i.e., by representing buildings with a basic flat roof. Recently new methods have emerged for producing models with a more detailed LOD2 CityGML formalism, tailored-made for the new generation of optical satellites (i.e., PLEIADES NEO) and the 30cm resolution of their images. Among the promising methods, the reconstruction pipelines based on the extrusion of a 2D wireframe describing the roof components from a near-Nadir image have shown a high potential. In particular, we demonstrated in prior work [2,3] that 2D wireframes can be extracted and lifted in 3D from PLEIADES NEO images with both a good accuracy and the geometric guarantees required for 3D applications. Unfortunately, the applicability of this promising work is limited to (i) a specific type of satellite data (i.e. PLEIADES NEO images), and (ii), a specific output representation (i.e. LOD2 CityGML formalism). In addition, this study exploits (iii) traditional Digital Surface Models (DSM) for 3D interpretation and largely ignores the emerging 3D representations based on radiance fields. 

Objectives

The objective of this PhD thesis is to design and implement a holistic pipeline for reconstructing urban scenes in 3D from satellite data. Departing from prior work on this topic [2], the PhD candidate will investigate the combination of geometric and learning models to generalize the wireframe extrusion concept to a wider range of input data (with a lower image resolution than in the PLEIADES NEO context) and of output 3D models (with a hierarchical multi-LOD representations). Three research directions will be explored.

Exploiting sub-metric satellite imagery. The PhD candidate will generalize the extrusion-based pipeline developed in [2] to a wider variety of satellites. One important issue will be to bring robustness to the reconstruction pipeline in presence of low-quality input image relatively to the PLEIADES NEO context. 

Generating Multi-LOD models. The PhD candidate will investigate reconstruction algorithms able to deliver 3D models of urban scenes at multiple LODs. If some methods [4,5] have reached this goal using plane assembly techniques from airborne data, the possibly to produce such polymorphic models from satellite data is a big scientific challenge which seems now affordable with the potential of the last computer vision foundation models. In particular, The PhD candidate will study how to combine building wireframe representations with foundation models able to analyze LOD3 details such as dormer-windows, chimneys and balconies. He/she will also investigate the multi-LOD reconstruction of trees and terrain.  

Exploiting emerging 3D representations. The candidate will also explore how emerging (radiance field-based) 3D representations such as Gaussian Splatting [6] can help improving the robustness of traditional geometric algorithms in challenging conditions, e.g., in the presence of buildings with reflective and transparent components [7]. He/she will investigate how traditional mesh processing, such as shape detection and mesh simplification can be adapted to these new representations originally made for solving image rendering problems.

References:

[1] Gröger, Kolbe, Nagel and Häfele. OGC City Geography Markup Language (CityGML) Encoding Standard, 2012

[2] Boyer. Geometric modeling of urban scenes with LOD2 formalism from satellite images. PhD thesis of University Côte d’Azur, 2025

[3] Boyer, Youssefi and Lafarge. LineFit: A Geometric Approach for Fitting Line Segments in Images. ECCV 2024

[4] Zhang, Pan, Lv, Gong and Huang. Architectural Co-LOD Generation. Trans. on Graphics, vol 43(6), 2024

[5] Verdie, Lafarge and Alliez. LOD generation for urban scenes. Trans. on Graphics, vol 34(3), 2015

[6] Kerbl et al. 3D Gaussian Splatting for Real-Time Radiance Field Rendering. Siggraph 2023

[7] Aira, Facciolo and Ehret. Gaussian Splatting for Efficient Satellite Image Photogrammetry. CVPR 2025

=================

For more Information about the topics and the co-financial partner (found by the lab!); contact Directeur de thèse - Florent.Lafarge@inria.fr

Then, prepare a resume, a recent transcript and a reference letter from your M2 supervisor/ engineering school director and you will be ready to apply online before March 13th, 2026 Midnight Paris time!