Microsoft Research India

Place Microsoft Research India , Hyderabad, India
Keywords Applied Computer Vision, Software Engineering
Role Research Engineer
Timeline Jul 2022 - Jul 2024

Highlights

  • IMWUT/ACM UbiComp 2023: HyWay: Enabling Mingling in the Hybrid World
  • Patent Filed: US Patent application filed in July 2023
  • Bill Gates Demo: Bill Gates personally demonstrated our project during his visit to Microsoft Research India
  • CEO Recognition: Our work was praised by Satya Nadella during the All Employees India Townhall in January 2023

HyWay: Hybrid Hallway

  • Primary contributor to HyWay - an interactive communication platform for in-person and remote participants.
  • Built an image editor powered by CosmosDB to create/edit event-spaces. Contributed to core-fixes such as memory-leakage: in-place mutation of game-objects, re-renders: migration from react state to redux store.

Indoor User Localization and Representation

  • Developed a computer-vision system to detect, track, localize and represent using multiple cameras.
  • Proposed an inter-camera matching algorithm to associate identical people across different camera viewpoints.
  • Optimized an open-source face detection lib (RetinaFace) and significantly reduced the end-to-end processing time of our system.
  • Real world evaluations showed our mean inter-camera matching algorithm accuracy is around 90%, mean indoor localization error is 1.037m and the overall pipeline achieved 15 FPS, supporting real-time operations.

Continual Learning for User State Modeling

  • Developed a heuristic + ResNet based continual learning framework for determining whether a user is distracted or not using multiple data sources in a privacy-preserving manner.
  • Implemented a data-processing module to collect and synchronize data from multiple sensors (e..g, cam, mic, keyboard, screen).
  • Implemented an acoustic echo cancellation module to facilitate usability of the tool in a headset-free setting.
  • Proposed a continual learning framework to generate annotated video data on fly and trained a ResNet to model each user.