Sajal Maheshwari

I am an Applied Scientist at Amazon Prime Video, where I work with the Virtual Product Placement team.

Virtual Product Placement(VPP) is a novel ad experience which to insert objects virtually within video shots in a photorealistic fashion. I am working on the occlusion handling thread of this highly complex problem at Amazon, using deep learning models to estimate an matting mask for occluded parts in a virtual insertion region. Previously, I have worked in the Autonomous Driving division at Qualcomm US with focus on object tracking. Before that, I have worked with the Camera team in Qualcomm India, and developed algorithms for better dynamic range and exposure. I graduated from Carnegie Mellon University, with a Masters degree in computer vision in December 2020 where I worked with Michael Kaess on Semantic Map Representation for indoor visual SLAM. I have completed my B.Tech. with Honours from International Institute of Information Technology, Hyderabad, in 2017, where I worked with Vineet Gandhi on document image quality analysis.

Email  /  CV  /  Bio  /  Google Scholar  /  Github

I'm interested in computer vision, machine learning, optimization, and image processing. I have worked in visual SLAM, document analytics and Computational Photography. Some of my published works are mentioned below

We try to predict the quality of an image by using the spatial frequency response function of the camera.

We try to predict the quality of an image by proposing a new method of ground truth generation in planar document images using spatial frequency response to generate local patch-based ground truths.

We estimate the amount of out-of-focus blur in a document image by creating a focal stack of images and assessing the quality of an image by the difference in the image and focus distance.

We estimate the amount of out-of-focus blur in a document image by extracting the important regions of an image by proposing a novel operator and determine the blur through statistical properties of the extracted regions.

We use classical image and gradient features along with pretrained VGG features to predict movement across specific regions in a video. Our insight is that using the combination of these specific regions and extracted features we can predict camera movement with higher accuracy(20%) and lesser memory constraints than current SOTA.

We use a novel combination of Kalman filter based tracking along with learning based object signatures created using re-identication models for efficient object tracking in autonomous driving systems.

Dean's Academic List and Dean's Research List, IIIT-H
Reviewer for ICVGIP 2018 and ICRA 2022

Website template credits