Saketh Rambhatla

I am a postdoctoral researcher at Meta AI working with Dr. Ishan Misra. I obtained my Ph.D. in ECE at University of Maryland (UMD), College Park, where I worked on in-the-wild visual understanding with Dr. Rama Chellappa and Dr. Abhinav Shrivastava.

I completed my bachelor's and master's at Indian Institute of Technology, Kharagpur. During my master's I worked on SLAM and pose recognition.

Email / CV / Google Scholar

Research

I'm interested in computer vision and deep learning. During my Ph.D. I worked on object tracking, person re-identification, object detection and discovery and multi-modal inconsistency detection tasks. Much of my research was on training models with improper supervision.

	MOST: Multiple Object localization with Self-supervised Transformers for object discovery. Sai Saketh Rambhatla, Ishan Misra, Rama Chellappa, Abhinav Shrivastava ICCV, 2023 (Oral Presentation) Project Page \| arXiv \| Code \| Poster Localize multiple objects in a real world images in an unsupervised fashion without any training using Self-supervised Transformers.
	SparseDet: Improving Sparsely Annotated Object Detection with Pseudo-positive Mining. Saksham Suri, Sai Saketh Rambhatla, Rama Chellappa, Abhinav Shrivastava ICCV, 2023 Project Page \| arXiv \| Code \| Poster Improve performance of detectors trained using missing annotations by posing the problem as region-based semi-supervised learning.
	The Pursuit of Knowledge: Discovering and Localizing new concepts using Dual Memory Sai Saketh Rambhatla, Rama Chellappa, Abhinav Shrivastava ICCV, 2021 Project Page \| arXiv \| Code \| Poster Equip machines with capabilities to automatically discover and learn models for new categories from a large unlabeled dataset.
	Towards Discovery and Attribution of Open-world GAN Generated Images. Saksham Suri, Sharath Girish, Sai Saketh Rambhatla, Abhinav Shrivastava ICCV, 2021 Project Page \| arXiv Automatically discover and attribute open-world GAN images.
	Self-Denoising neural networks for few shot learning Steven Schwarcz, Sai Saketh Rambhatla, Rama Chellappa arXiv Novel architecture based on denoising auto-encoders to improve few shot learning.
	An Empirical analysis of Boosting Deep Networks Sai Saketh Rambhatla, Michael Jones, Rama Chellappa IJCNN, 2022 Paper Empirical evidence that a single large neural network is usually more accurate than a boosted ensemble of neural networks with the same number of total parameters
	Towards Accurate Visual and Natural language-based vehicle retrieval systems. Pirazh Khorramshahi, Sai Saketh Rambhatla, Rama Chellappa NVIDIA AI City Challenge, CVPR Workshops, 2021 Paper Proposed a real-time system for image-based vehicle re-identification and natural language-based vehicle retrieval.
	Towards Real-Time Systems for Vehicle Re-Identification, Multi-Camera Tracking, and Anomaly Detection. Neehar Peri, Pirazh Khorramshahi, Sai Saketh Rambhatla, Vineet Shenoy, Saumya Rawat Jun-Cheng Chen, Rama Chellappa NVIDIA AI City Challenge, CVPR Workshops*, 2020 Paper Proposed a Real-Time system for Vehicle Re-identification, Multi-Camera Tracking, and Anomaly Detection in a network of traffic cameras.
	Detecting Human-Object Interactions via Functional Generalization. Ankan Bansal, Sai Saketh Rambhatla, Abhinav Shrivastava, Rama Chellappa AAAI, 2020 project page / arXiv Humans interact with functionally similar objects in a similar manner.
	Spatial Priming for Detecting Human-Object Interactions Ankan Bansal, Sai Saketh Rambhatla, Abhinav Shrivastava, Rama Chellappa arxiv project page / arXiv A method for exploiting the spatial layout information of a human and an object for detecting HOIs in images.
	A Dual-Path Model With Adaptive Attention for Vehicle Re-Identification Pirazh Khorramshahi, Amit Kumar Neehar Peri, Sai Saketh Rambhatla, Jun-Cheng Chen, Rama Chellappa ICCV, 2019 (Oral Presentation)* arXiv Proposed a novel dual-path adaptive attention model for Vehicle re-identification.
	Body Part Alignment and Temporal Attention Pooling for Video-Based Person ReIdentification Sai Saketh Rambhatla, Michael Jones BMVC, 2019 Paper Training deep networks with the ability to align features achieves state of the art performance on Person Re-identification.