Saksham Suri

I am a Ph.D student in the department of Computer Science at University of Maryland (UMD), College Park, advised by Prof. Abhinav Shrivastava. My research lies in the field of computer vision and machine learning.

During my PhD I have been fortunate to intern with Meta and Amazon. I completed my undergrad at IIIT Delhi (2019) majoring in Computer Science and Engineering where I worked at IAB Lab and Precog.

In the past, I have been fortunate to have worked with Rama Chellappa, Mayank Vatsa, Richa Singh, Ponnurangam Kumaraguru, C. -C. Jay Kuo, Anush Sankaran who have helped me grow as a researcher and a person.

Email  /  CV  /  Google Scholar  /  Twitter

I am interested in solving problems using less supervision and uncurated as well as synthetic data. Recently I have been working on improving recognition using generation especially using diffusion models as synthetic data sources. I have previously explored tasks across recognition and generation focusing on different supervision strategies and propose modified architectures and losses to utilize the data better under different settings.

Utilizing synthetic data from state-of-the-art diffusion models to improve object detection and segmentation performance.

Self-supervised and lightweight technique to learn a feature transform for generating dense features from pre-trained ViTs.

Decouple the optimization of reconstruction and adversarial losses by synthesizing an image as a combination of its reconstruction (low-frequency) and GAN residual (high-frequency) components.

We propose Diff2Lip an audio-conditioned diffusion-based model which is able to do lip synchronization in-the-wild while preserving image fidelity and identity.

Propose an end-to-end system that learns to separate the proposals into labeled and unlabeled regions using Pseudo-positive mining to tackle sparsely annotated object detection.

Study the effect of supervision and losses in trianing ViTs through attention, feature and downstream task based analysis.

Proposed an iterative algorithm for discovering images generated from GANs in an open world setup. Also show applications in an online never ending discovery.

We propose a novel framework which disentangles spatial and style information for image synthesis. A latent spatial layout for the target image is generated, which is used to produce per-pixel style modulation parameters for the final synthesis..

Incorporating task agnostic color, shape, texture and symmetry attributes to task specific deep learning classifiers for face recognition.

Propose a non deep learning based approach to handwritten digit synthesis which is more interpretable and does not require back-propogation.

Analyzing and identifying the presence of fake tech support accounts on twitter.

A novel approach to perform face recognition in the presence of plastic surgery and disguise.

Template credits