‪Roger Marí‬ Molas

Computer Vision Research Engineer (PhD) with a strong background in image processing and deep learning. My doctoral research focused on remote sensing and 3D vision tasks (reconstruction, calibration, co-registration, change detection). More recently, my work has centered on neural rendering for remote-sensing imagery, cultural heritage digitization, and the development and application of generative AI methods.

Born in Barcelona (1995), I studied at Universitat Pompeu Fabra, completing with honors a BSc in Audiovisual Systems Engineering and a specialized MSc in Computer Vision. I moved to Paris in October 2018 to pursue a PhD at Centre Borelli (ENS Paris-Saclay) under the supervision of Gabriele Facciolo. I defended my thesis Applications of multi-image remote sensing in December 2022. In January 2024 I joined Eurecat where I currently contribute to Computer Vision and AI projects in Catalonia and Europe.

Google Scholar  /  ResearchGate  /  LinkedIn  /  GitHub

I am passionate about scientific writing and communicating research through peer-reviewed publications. Selected publication highlights are listed below.

ShinyNeRF advances NeRF-based 3D digitization of specular surfaces by phisically modeling isotropic and anisotropic reflections based on interpretable material/geometric parameters (normals, tangents, ASG anisotropy), enabling anisotropic material editing.

This new dataset comprises multi-view satellite images (PAN, RGB), corresponding vegetation and shadow masks, bundle-adjusted RPC camera models and ground-truth DSMs for 702 different geographic areas of 500x500 m each across three different US cities.

We evaluate the fidelity and realism of different architectural variations of a latent diffusion model, which is used to generate RGB aerial images conditioned to semantic maps.

EO-NeRF is extended to handle high-res panchromatic (PAN) and low-res multispectral (MS) inputs, eliminating the need for separate pansharpening. The resulting model can render pansharpened image surrogates with high-res color information for each input viewpoint.

We propose a generic regularization framework for NeRF based on differential geometry that outperforms previous state-of-the-art methods with only three input views. We compare our approach with RegNeRF (CVPR 2022).

We present EO-NeRF, that reveals scene geometry from multi-date satellite images with an unprecedented level of detail. We propose a geometrically consistent shadow model and a radiometric decomposition of the scene adapted to pansharpened satellite images.

We evaluate the performance of the deep learning architectures PSM (CVPR 2018) and HSM (CVPR 2019) for disparity estimation on multiple pairs of high-resolution satellite images.

Sat-NeRF is the first work in neural rendering for multi-date satellite images to demonstrate quantitatively convincing results in terms of surface reconstruction.

We emulate a perfect sensor to generate a single image from a fragmented push-frame strip. The resulting product simplifies large-scale 3D modeling from push-frame imagery.

We propose a generic bundle adjustment method for multi-view stereo pipelines for satellite images. The RPC camera models of the input views are refined with a rotation that compensates localization errors related to the attitude angles encoding the satellite orientation.

The RPC camera models of a time series of SkySat acquisitions are refined and used to compute a surface model for each date, which is used to measure the stockpile volume.

We describe a terrain-independent algorithm to accurately derive the RPC camera model linking a set of 3D-2D point correspondences based on a regularized least squares fit.

This work investigates and compares different relative geolocation correction techniques for multi-view stereo pipelines for satellite images. We assess the impact on the output geometry.

We present a deep learning method to predict extrinsic (tilt and roll) and intrinsic (focal length and radial distortion) parameters from a single image. We use a parameterization that is better suited for learning than directly predicting the camera parameters.