Sareh Shirazi

About

Sareh Shirazi is a robotics and computer vision researcher whose work sits at the intersection of deep learning, visual place recognition, and autonomous systems. Best known for her landmark 2015 study "On the Performance of ConvNet Features for Place Recognition," which has accumulated over 530 citations, Shirazi systematically investigated how convolutional neural network features—transformative in classical computer vision—translate to the demanding conditions of robotic navigation and SLAM. Her findings exposed critical gaps between standard deep learning benchmarks and real-world robotic deployment, shaping how subsequent researchers approached visual localization. Complementing this, her work on sequence-based place recognition tackled the dual challenge of extreme appearance and viewpoint variation, pushing the boundaries of condition-invariant route following. Beyond localization, Shirazi has contributed to action recognition, real-time action prediction, and imitation learning, exploring how robots can learn tasks directly from visual demonstrations. Her early robustness analysis of Deep Q Networks also reflects a broader commitment to understanding the reliability of deep reinforcement learning agents. Across her career, Shirazi has consistently asked not just whether deep learning works, but under what conditions and limitations—a rigorous perspective that continues to inform the robotics research community.

Research Focus

Key Achievements

Top Papers

1. 1
   On the performance of ConvNet features for place recognition

   532 citations · 2015
2. 2
   On the Performance of ConvNet Features for Place Recognition

   135 citations · 2015
3. 3
   Sequence searching with deep-learnt depth for condition- and viewpoint-invariant route-based place recognition

   57 citations · 2015
4. 4
   A robustness analysis of Deep Q Networks

   5 citations · 2016
5. 5
   On Encoding Temporal Evolution for Real-time Action Prediction

   5 citations · 2017
6. 6
   What Would You Do? Acting by Learning to Predict

   4 citations · 2017
7. 7
   Enhancing human action recognition with region proposals

   2 citations · 2015
8. 8
   The need for dynamic and active datasets

   2 citations · 2015

Key Collaborators

Contact & Links