Anurag Ghosh

I'm an AI Researcher specializing in vision foundation models and robotics, passionate about building full-stack embodied systems, from novel algorithms published at CVPR, NeurIPS, and ICCV, to deployed systems impacting millions of users.

Currently, I'm a final year Robotics PhD at Carnegie Mellon, advised by Srinivasa Narasimhan. I work on open-world perception and planning. I was co-advised by Srinivasa Narasimhan and Christoph Mertz when I got my Masters here.

Earlier, I spent a few wonderful years at Microsoft Research working at the intersection of scalable perception and distributed systems. In a different era, I studied at IIIT Hyderabad and worked on structured perception from broadcast sports videos.

Here's my Resume.

Open to opportunities: I'm on the industry job market. If you think we'd be a good fit, let's chat!

Email GitHub Google Scholar LinkedIn

Research

I love building intelligent systems that work to delight millions and hopefully, billions of users. I'm broadly interested in closing the perception-action loop for robotics in our extremely messy and beautiful open physical world, while ensuring that the systems we build are efficient, frugal, accessible, and deployable at scale.

RAD-LAD: Rule and Language Grounded Autonomous Driving in Real-Time

Anurag Ghosh, Srinivasa Narasimhan, Manmohan Chandraker, Francesco Pittaluga

Website Paper Video

A real-time Language-Action planner combining grounded rule-based reasoning with Large Language Models (LLMs) to handle complex, long-tailed autonomous driving scenarios.

ROADWork Dataset: Learning to Recognize, Observe, Analyze and Drive Through Work Zones

Anurag Ghosh, Shen Zheng, Robert Tamburo, Khiem Vuong, Juan R. Alvarez Padilla, Hailiang Zhu, Michael Cardei, Nicholas Dunn, Christoph Mertz, Srinivasa Narasimhan

International Conference on Computer Vision (ICCV), 2025

Website Paper Code

A comprehensive open-source dataset and benchmark for studying long-tail, complex autonomous driving scenarios in work zones.

AerialMegaDepth: Learning Aerial-Ground Reconstruction and View Synthesis

Khiem Vuong, Anurag Ghosh, Deva Ramanan*, Srinivasa Narasimhan*, Shubham Tulsiani*

Conference on Computer Vision and Pattern Recognition (CVPR), 2025

Website Paper Code

A scalable data generation framework that combines mesh-renderings with real images, enabling robust 3D reconstruction across extreme viewpoint variations (e.g., aerial-ground).

Saliency Guided Image Warping for Unsupervised Domain Adaptation

Shen Zheng★, Anurag Ghosh★, Srinivasa Narasimhan

Winter Conference on Applications of Computer Vision (WACV), 2025

Website Paper Code

An unsupervised domain adaptation approach that oversamples salient regions via in-place image warping during self-distillation. It improves model robustness across diverse geographies, lighting, and weather conditions—facilitating reliable real-world deployment.

Learned Two-Plane Perspective Prior based Image Resampling for Efficient Object Detection

Anurag Ghosh, N Dinesh Reddy, Christoph Mertz, Srinivasa Narasimhan

Conference on Computer Vision and Pattern Recognition (CVPR), 2023

Website Paper Code

A learnable, geometry-aware image resampling approach that incorporates 3D scene priors (ground and sky planes) into the perception pipeline, significantly improving both the efficiency and accuracy of detecting small, far-away objects.

Chanakya: Learning Runtime Decisions for Adaptive Real-Time Perception

Anurag Ghosh, Vaibhav Balloli, Akshay Nambi, Aditya Singh, Tanuja Ganu

Conference on Neural Information Processing Systems (NeurIPS), 2023

Website Paper Code

Honorable Mention, Streaming Perception Challenge, CVPR 2021.

An RL-based control policy that observes scene and system state to make adaptive runtime decisions (model selection, resolution, compute allocation). It jointly optimizes accuracy and latency under strict real-time constraints for both server and on-device hardware.

REACT: Streaming Video Analytics On The Edge With Asynchronous Cloud Support

Anurag Ghosh, Srinivasan Iyengar, Stephen Lee, Anuj Rathore, Venkat Padmanabhan

International Conference on Internet of Things Design and Implementation (IoTDI), 2023

Paper

A fusion framework for video analytics that fuses on-device and cloud predictions asynchronously, improving object detection accuracy by up to 50% over device-only or cloud-only approaches.

Holistic Energy Awareness for Intelligent Drones

Srinivasan Iyengar, Ravi Raj Saxena, Joydeep Pal, Bhawana Chhaglani, Anurag Ghosh, Venkat Padmanabhan, Prabhakar T. Venkata

International Conference on Systems for Energy-Efficient Built Environments (BuildSys), 2021

Paper

Best Paper Runner-Up

A holistic energy management framework for intelligent drones that jointly optimizes compute, communication, and flight energy. Also appeared in Transactions on Sensor Networks.

Smartphone-based Driver License Testing

Watch Microsoft CEO Satya Nadella explain the project!

Read about our work on PM Awards Innovations Coffee Table Book! (Extracted here)

Deployed in multiple states/10+ cities in India, automatically testing hundreds of thousands of drivers at a low-cost with >99% accuracy (test verified by human operator). See Overview and Dashboard.

A deployed nationwide automated driver's license testing platform (100K+ users) replacing expensive pole-mounted infrastructure with a $500 smartphone. Implements robust monocular 3D localization (SfM/SLAM), IMU-based jerk recognition, and multi-modal perception for accurate trajectory estimation and driver state monitoring.

Relevant Publications

Smartphone-based Driver License Testing
Anurag Ghosh, Vijay Lingam, Ishit Mehta, Akshay Nambi, Venkat Padmanabhan, Satish Sangameswaran
Conference on Embedded Networked Sensor Systems (SenSys Demo), 2019

Paper

ALT: Towards Automating Driver License Testing using Smartphones
Akshay Nambi, Ishit Mehta, Anurag Ghosh, Vijay Lingam, Venkat Padmanabhan
Conference on Embedded Networked Sensor Systems (SenSys), 2019

Paper

Racket Sports Analysis — Federer, Nadal, Djokovic

Analyzing Racket Sports From Broadcast Videos

Anurag Ghosh

IIIT Hyderabad (Master's Thesis), 2019

Thesis

A geometry-aware perception system for broadcast racket sports. Automatically performs court-mapping, player tracking, and shot classification from monocular TV footage to quantitatively compare the playing styles of Federer, Nadal, and Djokovic.

Towards Structured Analysis of Broadcast Badminton Videos

Anurag Ghosh, Suriya Singh, C.V. Jawahar

Winter Conference On Applications of Computer Vision (WACV), 2018

Paper Real-Time Deployment

Piloted with ESPN/Star Sports at Premier Badminton League, watched by tens of millions in South East Asia.

An end-to-end framework for real-time player analysis from live broadcast badminton videos using only visual cues — computing on-court distance, speed, and heatmaps without multi-modal sensors.