Context

Reconstructing 3D scenes from multiple camera views—essentially inverting ray tracing—is a cornerstone of computer vision. Since the 1970s this problem has steadily advanced, reaching a breakthrough in 2020 with NeRF (Neural Radiance Fields) [Mildenhall et al., 2021], which uses AI optimization tools to invert the image formation model and delivers unprecedented visual fidelity from several optical views. Its flexibility also enables fusion of diverse observation modalities, and it has already proven effective for remote sensing applications [Derksen & Izzo, 2021; Marí et al., 2023]. This is a crucial problem in remote sensing and satellite imaging, as Earth observation satellites offer a large variety of active and passive views of the ground, be it by optical or synthetic aperture radar (SAR) satellites, obtained at different dates.

Since NeRF, other techniques such as Signed Distance Functions (SDF) [Yariv et al. 2021] and Gaussian Splatting (GS) [Kerb et al. 2023] have been proposed, offering alternative ways of encoding scene geometry. They present different trade-offs in efficiency and flexibility, that must be investigated in the remote sensing context [Aira et al. 2025]. Gaussian Splatting is particularly interesting as it represent a substantial speedup with respect to the other techniques. Furthermore, in [Ehret et al. 2024b] it was shown that inverse rendering techniques can be applied to SAR imagery to recover 3D surface geometry from non-interferometric SAR images. This proof of concept, though closer to SDF [Yariv et al. 2021], highlights the potential of inverse rendering approaches for SAR imagery.

This thesis aims to explore the extension of Gaussian Splatting, to radar imagery. This research seeks to adapt and optimize GS for handling the unique characteristics of radar data: speckle noise, acquisition geometry, corner reflectors, and occlusion handling. The ultimate goal is to fuse observation from optical and SAR modalities in a single inverse model. 

Methodology

The first problem to be addressed in the thesis is the radar image formation within the GS framework. Some recent works [Li et al. 2025][Liu et al. 2024] have proposed extensions of GS to radar, however they seem to be specialized for vehicle reconstruction from aerial imagery. 

Secondly, these extensions, although very relevant, leverage spherical harmonics of each Gaussian to model the signal intensity. This is a key difference with respect to Radar Fields [Ehret et al. 2024b] that computes the intensity from the surface normals. We thus propose to compare these methods with techniques that allow to associate a normal orientation to the GS reconstruction. Some recently proposed variants of GS [Chen et al. 2024][Guédon et al. 2025] use flattened gaussians to represent the scene, or use the gaussians as proxy for an underlying flat support. 

Planning 

- The thesis will start with a literature review on of existing GS methods and their applications in optical imagery, alongside a survey of current approaches for 3D reconstruction from SAR data.

- The development of the GS framework for SAR images will rely on simulated data, therefore a simple simulator will be developed as a way to generate synthetic data for the experiments, but especially in order to assimilate all the needed SAR concepts. 

- Replicating the results from [Li et al. 2025][Liu et al. 2024] will be a first step to address the problem of improving the inverse rendering model. Special attention will be put on the handling of corner reflectors and occlusions, and extensions will be proposed accordingly.

- Further extend the GS formation model by computing the intensity using the surface normal (following [Chen et al. 2024][Guédon et al. 2025] or similar works).  Estimating surface normals will likely require imposing explicit regularization in the solution [Ehret et al. 2024a]. 

- Validate the proposed framework using both simulated and real-world SAR datasets, establishing benchmarks against current techniques as Radar Fields [Ehret et al. 2024b].

- The last phase of the thesis will be to fuse observation from optical (e.g. [Aira et al. 2025]) and SAR modalities in a single inverse model and study the gain contributed by each modality.

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

For more Information about the topics and the co-financial partner (found by the lab!); contact Directeur de thèse - gabriele.facciolo@ens-paris-saclay.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!