Toward Robust Neural Reconstruction from Sparse Point Sets
A distributionally robust optimization (DRO) framework stable and efficient optimization of SDFs without ground truth supervision
I am a PhD Student at INRIA in France, working with Prof. Adnane Boukhayma and Prof. Eric Marchand.
My research interests focus on 3D computer vision, generative models, and computer graphics. Previously, I worked on improving generalizable implicit reconstuction models. I’m currently exploring implicit representation learning from sparse inputs (pointclouds, multi-view images).
A distributionally robust optimization (DRO) framework stable and efficient optimization of SDFs without ground truth supervision
A new theoretical understanding of training-free few-shot adaptation for CLIP
Imporving Novel-View Synthesis and reconstruction from sparse views without data priors through local Geometric Linearization.
Local linearization and Minimal entropy fields regularization for occupancy learning from sparse point cloud without ground-truth supervision.
Adversarial regularization for SDF optimization from sparse point cloud without ground-truth supervision.
Making generalizable implicit representations robust to the input noise and sparsity by relaxing the inductive biais of convolutions with MLP-based local encoding.
Leveraging kernel methods for Fast test-time adaptation of pretrained networks with a shape specific trade-of between expressiveness and regularization.
Conditioning implicit representations on sparse inputs through meta-leanring in feature space.