Rajagopalan N Ambasamduram

Indian Institute of Technology Madras
1 previous affiliation
3Machine learning papers tracked by Lacuna
Is this you?

Research Overview

Rajagopalan N Ambasamduram focuses on computational photography, image deblurring, and implicit neural representations for video data. Early collaborative work addressed non-uniform motion blur in high dynamic range imaging and dual-lens camera systems. More recently, the research trajectory has shifted toward developing efficient spatiotemporal neural architectures for continuous video representation.

Research Trajectory

The available evidence outlines a research trajectory spanning from 2013 to 2023, primarily centered on computational photography, image restoration, and video representation. Early work in this corpus demonstrates an interest in the complex physical modeling of camera motion and exposure. In 2013, Ambasamduram participated as a co-author in the development of Non-Uniform Deblurring in HDR Image Reconstruction. This research addressed the compounding problem of camera shake in high dynamic range (HDR) imaging, where capturing multiple exposures increases the likelihood of blur. The work utilized transformation spread functions to model the non-uniform blur resulting from both rotational and translational camera motion.

Moving toward the end of the decade, the focus shifted to accommodate emerging hardware paradigms, specifically the proliferation of multi-camera systems in mobile devices. As senior author, Ambasamduram guided work on Unconstrained Motion Deblurring for Dual-Lens Cameras (2019). This research introduced a generalized blur model designed to explain the intrinsically coupled image formation process in dual-lens setups, providing a dedicated deblurring methodology for unconstrained camera arrays.

Most recently, the trajectory has evolved from classical physical blur modeling to deep learning-based video representation. In 2023, serving again as senior author, Ambasamduram oversaw the publication of PNeRV: A Polynomial Neural Representation for Videos. This work tackled the challenges of extracting Implicit Neural Representations (INRs) for video data, a domain where traditional frame-only parameterizations often sacrifice spatiotemporal continuity. By leveraging Polynomial Neural Networks, the research introduced a parameter-efficient, patch-wise INR that modulates continuous spatial signals with continuous time signals to preserve continuity across both dimensions.

Current Directions

Based on the most recent evidence from 2023, current research directions center on advanced neural architectures for video processing and representation. A primary focus is the development of Implicit Neural Representations that can effectively handle the temporal dimension of video data without sacrificing spatial continuity. This involves exploring polynomial neural networks and continuous signal modulation to create parameter-efficient, spatiotemporal models. The shift from traditional computational photography techniques to implicit neural modeling suggests an ongoing interest in optimizing how dynamic visual data is parameterized, compressed, and reconstructed within modern deep learning frameworks.

Profile updated

See something wrong? Report a mistake