Portrait of Deepti Rawat

Deepti Rawat

PhD student · CVIT, IIIT Hyderabad

I'm a PhD student in Computer Science at IIIT Hyderabad, working with Prof. Ravi Kiran Sarvadevabhatla in the Center for Visual Information Technology (CVIT). My research focuses on addressing safety challenges in autonomous systems, particularly in unstructured navigation scenarios.

My work leverages perceptual cues to develop learning-based solutions for real-world scenarios. This includes traffic violation detection for two-wheelers, context-aware modeling of riders and motorcycles in occluded traffic scenarios, and designing framework to probe Vision LLMs' capabilities in understanding diverse road conditions. Broadly, my goal is to achieve trustable autonomous driving systems.

Outside of research, I enjoy running, reading, and spending time with friends.

Publications & news

Publications

Ask the benchmark about this clip:

Benchmark answer Pick a question. These are real QA pairs from RoadSocial for this exact video.

RoadSocial: A Diverse VideoQA Dataset and Benchmark for Road Event Understanding from Social Video Narratives

Chirag Parikh†, Deepti Rawat†, Rakshitha R. T., Tathagata Ghosh, Ravi Kiran Sarvadevabhatla  ·  †equal contribution

CVPR 2025
Problem
Understanding a road event takes more than seeing it. It takes knowing where you are, what's normal there, and what nearly happened but didn't.
Approach
We convert social commentary on road videos into video question-answer supervision, drawing on contextual insight the footage alone cannot supply.
Result
A benchmark that exposes where Video LLMs actually break (timing, reasoning, and hallucination), and points to social narrative as a promising source of supervision for the parts models find hardest.
Project arXiv Dataset Code
Live system output, pick a violation:

E-ticket · triple riding Three riders detected, associated with their motorcycle through occlusion, tracked, plate read.

DashCop: Automated E-Ticket Generation for Two-Wheeler Traffic Violations using Dashcam Videos

Deepti Rawat†, Keshav Gupta†, Aryamaan Basu Roy, Ravi Kiran Sarvadevabhatla  ·  †equal contribution

WACV 2025 Deployed in production
Problem
Enforcement only works where it can see. Fixed cameras watch a handful of junctions, while the violations happen everywhere else.
Approach
We move enforcement onto the dashcam. DashCop reads ordinary driving footage end to end, learning which riders belong to which motorcycle, holding that grouping together through traffic and occlusion, and carrying it through to a license plate and a ticket.
Result
A system evaluated end to end, at the level of the ticket rather than its parts.

Running on real roads: powers VIOLA, deployed with the Hyderabad Traffic Police.

Project arXiv Dataset Code
A two-wheeler rider's view of Hyderabad traffic: a scooter rider ahead, an auto-rickshaw, and oncoming vehicles
Where did the rider look? Take a guess:

Try it Click anywhere on the frame. The rider's recorded gaze point appears in red.

myEye2Wheeler: A Two-Wheeler Indian Driver Real-World Eye-Tracking Dataset

Bhaiya Vaibhaw Kumar, Deepti Rawat, Tanvi Kandalla, Aarnav Nagariya, Kavita Vemuri

ITSC 2024
Problem
Driver attention models are built from four-wheeler drivers on well-planned Western roads. A two-wheeler rider in dense mixed traffic sees the road differently.
Approach
We put an eye tracker under the helmet and sent riders onto a real Indian road, recording where they looked as they negotiated live unstructured traffic.
Result
A saliency model that reads European four-wheeler gaze well falls off sharply here. Attention on these roads follows patterns a model trained elsewhere has not learned.
arXiv Dataset

News