Mukund Varma T

I'm a first year CSE Ph.D. student at UC San Diego, jointly advised by Professors Ravi Ramamoorthi and Hao Su. I received my bachelors degree from IIT Madras.

I was fortunate to have spent several summers at Adobe, Nvidia, and Google with amazing mentors and colleagues. I have also had the opporunity to collaborate (and still continue to) with folks from the VITA Group at UT Austin. During my undergrad, I was also associated with the Advanced Geomtric Computing Lab, and the Computational Imaging Lab.

I am lucky to be supported by the Jacobs School of Engineering Fellowship.

Email  /  CV  /  Scholar  /  Twitter  /  Github

My current research interests lie at the intersection of computer vision, computer graphics and machine learning, specifically to facilitate high quality 3D reconstructions and semantic understanding from multiple view points. Some papers are highlighted.

A generalizable framework for novel view synthesis using degraded input captures containing any imperfection type.

A camera-free novel view synthesis technique from sparse input views (as few as 3 images of large-scale scenes).

General framework to lift any pretrained 2D vision model to generate 3D consistent outputs with no additional optimization.

Improves consistency of off-the-shelf multi-view generation techniques using a generalizable SDF, enabling super fast 3D generation from a single image.

A unsupervised learning pipeline for generalizable novel view synthesis and restoration of underwater scenes by disentangling into individual image formation components.

We scale up generalizable NeRF training by borrowing the concept of mixture of experts from language models.

We present emperical evidence that convolutional networks trained on sift features improve robustness to unseen out-of-domain data with minimal to no-loss in in-domain performance.

We propose a generalizable neural scene representation and rendering pipeline that achieves superior quality compared to previous methods.

Sparse networks identified using iterative magnitude pruning showcase improved data-efficiency and robustness compared to their dense counterparts.

Collaborative repository of natural language transformations.

Collaborative benchmark for measuring and extrapolating the capabilities of language models.

A specialized pipeline for point cloud shape completion that can generalize to synthetic and real partial scans from seen and unseen categorical types.

We introduce Hierarchical Attention, a reccurent transformer module that immitates convolution-like operation with significantly lower computational budget.

We propose a transformer architecture for sparse set learning, e.g. point cloud understanding.

• Conference Reviewer:
  NeurIPS (2022, 23, 24), ICLR (2023, 24), ICCV (2023), ECCV (2024), CVPR (2024).
• Journal Reviewer:
  Computers and Graphics (2021, 22).
• Teaching Assistant (IIT Madras):
  
  • CS6700 Reinforcement Learning, Spring 2023.
  • EE5180 Introduction to Machine Learning, Spring 2023.
  • EE5179 Deep Learning for Imaging, Fall 2022.

I enjoy hiking ⛰️ (cuz duh I am a boring computer science kid), used to sketch ✏️ a bit, and love playing most sports, particularly badminton 🏸, tennis 🎾 and soccer ⚽ (ahem I am american XD). Feel free to reach out to chat about anything, but headsup I love to yap (it can be scary, umm I am kidding lmao). Fun fact: my name in chinese 牧坤 (Mù kūn) when broken down essentially means "to manage" 牧 (Mù) and 坤 (kūn), one of the eight trigrams in I Ching that represents "the earth" 😤.

Template stolen from here.