I am Latha Pemula, I am known for my work in Anomaly Detection. Over the years that I worked in anomaly detection, I made a few, but meaningful and impactful contributions to the field.  I have built PatchCore, the anomaly detection method that stood test of time by still being the benchmark 5 years after its release. PatchCore is deployed for many different applications across many different industries across the world.  I have steered the academic  community to work on problems that are relevant to practical applications by setting the benchmarks  - VisA dataset is now widely used benchmark for visual anomaly detection and the recently released RobustAD benchmark sheds light on the domain generalization problem in Anomaly detection and spurred new research.

Looking Forward ..

I believe in research that stands the test of time and makes meaningful albeit small steps. I believe my best work is yet to come.

I am looking for opportunities in Multimodal AI, World Models, Robotics and perception.

Past Experience

I worked as an Applied Scientist at AWS AI Labs for 10 years. I am one of the founding team members responsible for launching AWS Lookout for Vision service. I have also part of the team that launched AWS Bedrock Data Automation service where I led a logo detection service and Contributed to Document Understanding modules. My research interests include Anomaly Detection, Fewshot learning, Representation Learning, Domain Generalization. Before Amazon, I completed my Masters at Virginia Tech under the guidance of Dr. Dhruv Batra and Dr. Devi Parikh. I had a chance to work under the guidance of Dr. Lei Zhang during my internship at Microsoft Research. My masters thesis and my internship focused on fewshot classification. I completed my Bachelors in Electrical Engineering at Indian Institute of Technology, Madras(IITM) in 2010.

Contact

email: latharampemula@gmail.com

Linkedin

Google Scholar

Semantic Scholar

Spot-the-Difference Self-Supervised Pre-training for Anomaly Detection and Segmentation

 We release the Visual Anomaly(VisA) Dataset consisting of 10,821 high-resolution color images (9,621 normal and 1,200 anomalous samples) covering 12 objects in 3 domains, making it the largest industrial anomaly detection  dataset to date. We also introduce a new self-supervised  learning method for  ImageNet  pre-training to   improve anomaly  detection and segmentation in 1-class and 2-class  5/10/high shot training  setups. 

ECCV, 2022

arXiv | code