We consider the challenging problem of learning Signed Distance Functions (SDF) from sparse and noisy 3D point clouds. In contrast to recent methods that depend on smoothness priors, our method, rooted in a distributionally robust optimization (DRO) framework, incorporates a regularization term that leverages samples from the uncertainty regions of the model to improve the learned SDFs. Thanks to tractable dual formulations, we show that this framework enables a stable and efficient optimization of SDFs in the absence of ground truth supervision. Using a variety of synthetic and real data evaluations from different modalities, we show that our DRO based learning framework can improve SDF learning with respect to baselines and the state-of-the-art.
Visual comparisons on the sparse reconstruction setting of Faust Dataset.
Visual comparisons of Reconstructions from VGGSfM point clouds of sparse views from BlendedMVS and Tanks & Temples datasets.
Visual comparisons of SemanticPOSS reconstructions from road scene LiDAR data.
@article{sdro_sdf,
title={Toward Robust Neural Reconstruction from Sparse Point Sets},
author={Ouasfi, Amine and Jena, Shubendhu and Marchand, Eric and Boukhayma, Adnane},
journal={arXiv preprint arXiv:2408.14724},
year={2024}
}