Mehar Khurana

Hi! I'm an MS in Robotics student at Carnegie Mellon University, advised by Prof. Deva Ramanan. I am currently working on 3D semantic representation learning.

I recently completed my undergrad at IIIT Delhi, where I worked with Prof. Saket Anand and Depanshu Sani on active learning in animal re-identification.

Previously, I was a research intern at the Toronto Computation Imaging Group (TCIG), where I worked with Prof. David Lindell and Anagh Malik on multi-view reconstruction using single-photon images, and curated a single-photon dataset.

Email  /  CV  /  Scholar  /  GitHub

Research

My research primarily focuses on understanding and interpreting the physical world.

Extending 2.5D feed-forward geometry predictors to "2.6D" by replacing per-pixel depth supervision with differentiable 3D Gaussian splatting, enabling supervision from novel views.

An active learning framework that uses disagreement between clustering methods to select informative and ambiguous samples, improving performance of animal re-identification models while minimizing annotation effort.

Pre-training on pseudolabels generated by 2D foundation models significantly improves 3D object detection accuracy with limited labeled data for both LiDAR-only and multi-modal detectors.

Source code stolen from Jon Barron's website.