Anisha Jain

I am currently pursuing Masters in Computer Vision (MSCV) at Robotics Institute at Carnegie Mellon University, advised by Prof. Lazlo A. Jeni and Michael Zollhoefer, Meta Reality Labs, where I am working on dynamic 3D reconstruction focusing non-rigidly deforming scenes captured by RGBD sensors. My primary interests lie in the intersection of Computer Vision and Machine Learning.

Prior to joining CMU, I worked as a Software Engineer at Microsoft (R&D) Pvt. Ltd., where my work focused on building agile deployment and experimentation of ML models for spam/phish detection during mail flow for Microsoft Defender for Office (MDO). I had the previledge of working with a team of talented engineers and researchers, and was mentored by Jay Goyal, Saurabh Shrivastav and Ganesh Pande.

In 2018, I worked under the supervision of Dr. Amarjot Singh and Dr S.N Omkar in the Computational Intelligence lab at Indian Institute of Science, Bangalore suspicious activity detection in surveillance using gait-based gender recognition system for subjects in loosely fitted clothing.

I completed my undergraduate education with a Major in Computer Science and Engineering from National Institute of Technology, Warangal in 2021. I had the pleasure of working under Dr. R Padmavathy for my undergraduate thesis.

In the late summer of 2020, I was a part of the Software Product Sprint at Google APAC, where I worked under the mentorship of Yash Jain. I have had the opportunity to intern at Microsoft Teams Mobile team, and The Garage at Microsoft (R&D) Pvt. Ltd., in the summer of 2020 and 2019 respectively.

Email  |  CV  |  Linkedin

Zoox
Computer Vision Intern
May 2024 - Aug 2024

Carnegie Mellon University
MS in Computer Vision
Aug 2023 - Dec 2024

Microsoft (R&D) Pvt. Ltd.
Software Engineer
June 2021 - Aug 2023
SWE Intern
May 2020 - July 2020
SWE Intern
May 2019 - July 2019

National Institute of Technology, Warangal
B.Tech. Computer Science
Aug 2017 - May 2021

Google APAC
Software Product Sprint
July 2021 - Sept 2021

Indian Institute of Science, Bengaluru
Research Intern
May 2018 - June 2019

Instead of using complex camera systems, the goal is to simplify avatar creation by only using smartphone data with rgb-d sensors which can profoundly enhance the scalability and adoption of avatars. We propose a novel method for dynamic 3D reconstruction focusing on non-rigidly deforming scenes captured by RGBD sensors.

We streamline 3D asset creation by editing existing scenes, reducing complexity and time. While NeRF offers detailed rendering, it's hard to manipulate. Instead, 3D Gaussian Splatting enables real-time rendering, simplifying tasks. Our architecture integrates vision language, allowing language-driven editing of visual properties. This bridges the gap between language understanding and 3D scene editing, enhancing workflows.

Our methodology can work with any text-to-image (T2I) diffusion architectures pre-trained on large image dataset. We introduce a spatiotemporal attention and background regularization techniques along with one-shot tuning strategy within our "GIF-Tune" model for temporally coherent and depth-consistent GIFs.

Alpha-MEPOL method for unsupervised RL in multiple environments achieves enhanced performance with dynamic alpha and a higher KL-Divergence threshold. Curiosity-driven exploration proves effective in high-dimensional environments, fostering diverse experiences and improving learning outcomes. Our experiments confirm positive results, highlighting the effectiveness of these modifications in the current context.

Addressing the challenge of identifying individuals wearing loosely fitted clothes of the opposite gender to evade detection, our proposed Bayesian Gait-based Gender Identification (BGGI) excels in real-world scenarios. Utilizing the LFCI dataset for training, our technique demonstrates superior performance in both pose estimation and gender recognition, employing state-of-the-art methods on dense real-world videos.

Source code from Jon Barron