Mateusz Michalkiewicz

I am a Postdoctoral Associate in the Electrical and Computer Engineering (ECE) department at Rice University (Houston, Texas, USA), working with Guha Balakrishnan. I received my PhD degree from The University of Queensland (Brisbane, Australia) supervised by Anders Eriksson and Mahsa Baktashmotlagh. During my PhD, I interned with the computer vision groups at Intel (group of Vladlen Koltun) and NEC-Labs (group of Manmohan Chandraker).

Google Scholar  /  DBLP  / 

My research interests broadly lie in deep learning and computer vision. Currently, I focus on reliability and geometric reasoning of vision and vision–language foundation models. I develop benchmarks and probes that stress‑test 3D understanding, symmetry perception, and invariance/equivariance to viewpoint changes.

We introduce GIQ, a comprehensive benchmark designed to evaluate the geometric reasoning capabilities of vision and vision-language foundation models.

We propose a systematic approach for evaluating the viewpoint stability of foundational models and its impact on 3D reasoning tasks.

We propose a novel view synthesis method for multi-elevation 3D building reconstruction with unknown camera poses

We introduce ShapeNetMini, a small subset of ShapeNet aimed at accelerating model training without sacrificing performance. Additionally, we enhance the performance of existing few-shot 3D reconstruction methods.

We propose a dropout method leveraging the gradient-signal-to-noise ratio (GSNR) of network's parameters to address overfitting, achieving competitive results on standard domain generalization benchmarks

We introduce three methods to learn class-specific global shape priors directly from data, tailored for few-shot 3D reconstruction from single images.

Our work demonstrates that high-quality 3D reconstruction can be achieved using a linear decoder derived from principal component analysis on the signed distance function (SDF) of the surface, allowing for easy scalability to larger resolutions

We demonstrate that defining individual layers in a neural network implicitly provides much richer representations over the standard explicit one, consequently enabling a vastly broader class of end-to-end trainable architectures.

We propose representing 3D shapes implicitly as oriented level sets using a continuous and discretized embedding function within a deep learning framework.

We propose a method for measuring brain atrophy, significantly reducing sample size estimates compared to state-of-the-art techniques

Website credits to Jon Barron