Soroush Saryazdi

I currently lead the Neural Networks team at Matic, supervised by Navneet Dalal. We are building robots that save people time and energy.

My interests are:
1. Building easy to use products which solve real problems
2. Designing perception systems
3. Developing accurate & reproducible on-device vision Neural Networks

I did my Computer Science Master's studies at Concordia University in collaboration with Liam Paull's Robotics and Embodied AI Lab (REAL) at Mila. I worked on using end-to-end trainable models for recovering the 3D geometry and rendering properties of a scene.

I live in the San Francisco Bay Area.

I sometimes review for NeurIPS, ICLR & ICML.

Email  /  Twitter  /  Google Scholar  /  Github  /  LinkedIn  /  CV

profile photo
News
May 2023 ConceptFusion accepted for demo at CVPR 2023
May 2023 ConceptFusion accepted at ICRA 2023 PT4R Workshop
April 2023 ConceptFusion accepted at RSS 2023
April 2023 Matician first product launch
April 2022 Highlighted reviewer for ICLR 2022
May 2021 Joined Matician as employee #30
May 2021 Defended my Master's thesis & graduated
April 2021 Our paper titled "Disentangled Rendering Loss for Supervised Material Property Recovery" won the Best Student Paper Award at GRAPP 2021.
Nov.
2020
We released gradslam: an open-source library for differentiable dense SLAM.
Featured Papers

Please see my Google Scholar for a complete list of papers.

ConceptFusion: Open-set Multimodal 3D Mapping
Krishna Murthy Jatavallabhula, Alihusein Kuwajerwala*, Qiao Gu*, Mohd Omama*, Tao Chen, Shuang Li, Ganesh Iyer, Soroush Saryazdi, Nikhil Keetha, Ayush Tewari, Joshua B. Tenenbaum, Celso Miguel de Melo, Madhava Krishna, Liam Paull, Florian Shkurti, Antonio Torralba
RSS 2023
CVPR 2023 Demo
ICRA 2023 PT4R Workshop
project page / paper / cite

ConceptFusion builds open-set 3D maps that can be queried via text, click, image, or audio.

Disentangled Rendering Loss for Supervised Material Property Recovery
Soroush Saryazdi, Christian Murphy, Sudhir P. Mudur
GRAPP 2021 (Best Student Paper Award)
paper / cite

The disentangled rendering loss can be used in supervised material appearance estimation tasks for recovering more accurate individual BRDF parameters.

SLAM: Automagically differentiable SLAM
Krishna Murthy Jatavallabhula*, Soroush Saryazdi*, Ganesh Iyer, Liam Paull
CVPR workshop 2020
project page / code / paper / cite

SLAM (gradSLAM) is a differentiable computational graph view of dense SLAM which allows gradients from SLAM outputs to be backpropagated to the input.

The Problem of Entangled Material Properties in SVBRDF Recovery
Soroush Saryazdi, Christian Murphy, Sudhir P. Mudur
Eurographics workshop 2020
paper / cite

We analyze the landscape of the rendering loss and how it can negatively impact the network into estimating less accurate individual BRDF parameters.


Template source code from Jon Barron's website