Srinath Ravi

Resume

About Me

I am an AI Research Scientist at UnscriptAI, where I focus on realistic talking head synthesis using 3D vision models. Previously, I was part of the Vision and AI Lab at IISc, led by Prof. Venkatesh Babu, and collaborated with Prof. Srinath Sridhar from Brown University and the Data Labs team at Capital One. My research primarily involves recent 3D representations, including Gaussian Splatting and NeRF.

News

Oct 2024
Our work, ChromaDistill, was accepted at WACV 2025!
Oct 2024
Joined UnscriptAI as an AI Research Scientist.
May 2024
Started working as a 3D-Vision Consultant with Capital One in collaboration with IISc.

Dec 2023
Invited as a speaker for the 4th edition of the Capital One ML Summit.

Nov 2023
Presented a poster at the Adobe-IISc GenAI Workshop.
Aug 2023
Our work, CoRF, was accepted at the "Second Workshop on AI for 3D Content Creation," ICCV 2023!
Jul 2023
Our work, Strata-NeRF, was accepted at ICCV 2023!
Aug 2022
Joined the Vision and AI Lab at Indian Institute of Science!
May 2022
Joined Springworks full-time as an SDE-1 (Machine Learning).
Mar 2022
Awarded "Intern of the Month" at Springworks.
Aug 2021
Started as a Machine Learning Engineer Intern at Springworks.
Mar 2021
Started as a Machine Learning Engineer Intern at ResoluteAI.in.
Aug 2018
Started B.Tech in CSE at DSCE, Bangalore.

Publications

Turbo-GS

Turbo-GS: Accelerating 3D Gaussian Fitting for High-Quality Radiance Fields

Tao Lu, * Ankit Dhiman, * Srinath R, * Emre Arslan, Angela Xing, Yuanbo Xiangli, R Venkatesh Babu, Srinath Sridhar

Under Review

Novel-view synthesis is key for 3D reconstruction and mixed reality, with 3D Gaussian Splatting (3DGS) offering real-time, high-quality results. However, 3DGS is slow to train, taking up to 30 minutes for 200 views. Turbo-GS speeds up optimization by using fewer steps, improving densification with error guidance, and introducing a convergence-aware budget control mechanism, all while maintaining high-quality rendering.
Project Page Paper Code
Acc3DSeg

Acc3DSeg: Accelerated 3D Segmentation via Contrastive Learning

Ankit Dhiman, * Srinath R, * Jaswanth Reddy, R Venkatesh Babu

Under Review

Recent advances in novel-view synthesis, including NeRF and 3DGS, have driven efforts to lift 2D segmentation labels to 3D. Existing methods often struggle with multi-view consistency and rely on costly two-stage processes. This work proposes an efficient, single-stage framework for 3D segmentation that eliminates offline clustering and object count dependencies, improving performance on complex real-world scenes.

Publication 3

ChromaDistill: Colorizing Monochrome Radiance Fields with Knowledge Distillation

Ankit Dhiman, Srinath R, Srinjay Sarkar, Lokesh R Boregowda, R Venkatesh Babu

WACV 2025 (Also appeared in AI3DCC workshop in ICCV'23)

This work presents a method for colorizing 3D scenes from grayscale multi-view images using pre-trained image colorization models. The approach ensures high-quality, consistent colorization without added computational overhead during inference. It is effective for NeRF and 3DGS representations, producing high-quality results across diverse scenes.

Project Page Paper Code
Publication 4

Strata-NeRF: Neural Radiance Fields for Stratified Scenes

Ankit Dhiman, Srinath R, Harsh Rangwani, Rishubh Parihar, Lokesh R Boregowda, Srinath Sridhar, R Venkatesh Babu

ICCV 2023

Strata-NeRF extends Neural Radiance Fields (NeRF) to model layered scenes, enabling smooth transitions between different scene levels. By conditioning on Vector Quantized (VQ) latent representations, it handles abrupt structural changes. Evaluations show it captures stratified scenes with high fidelity and minimal artifacts, outperforming existing methods.

Project Page Paper Code

Experience

UnscriptAI

UnscriptAI

AI Research Scientist — to

  • Currently developing Talking Head synthesis models leveraging 3D computer vision techniques.
Capital One

Capital One

Consultant — to

  • Developed an end-to-end immersive experience pipeline for Capital One’s vehicle reselling vertical, Auto Navigator, in collaboration with IISc.
  • Implemented methods to detect inconsistencies in COLMAP by analyzing extrinsic camera information, reducing computation costs and flagging inconsistencies in real-world mobile phone captures.
  • Leveraged various NeRF and Gaussian Splatting techniques to reconstruct vehicle models under diverse lighting conditions and camera types.
Vision and AI Lab

Vision and AI Lab, IISc

Research Assistant — to

  • Designed and developed Strata‑NeRF, an advanced Implicit NeRF model utilizing VQ‑VAE to handle stratified scenes. This model surpassed state-of-the-art techniques such as MipNeRF‑360, TensoRF, and InstantNGP and was accepted at ICCV 2023.
  • Contributed to the creation of ChromaDistill, a Radiance Field Network capable of colorizing grayscale and infrared images with 3D consistency. This work was accepted at WACV 2025.
  • Developed Turbo‑GS, a cutting-edge Gaussian fitting algorithm that performed efficient optimization, achieving significantly faster convergence with reduced memory usage.
Springworks

Springworks

SDE-1 (Machine Learning) — to

Machine Learning Engineer Intern — to

  • Enhanced document parsing efficiency by implementing a solution with PDFMiner, reducing reliance on external APIs and decreasing API call usage by 30%, thereby optimizing operational costs.
  • Contributed to the development of SpringVerify, a key background verification tool, by extracting critical data from various document types, including driving licenses, PAN cards, and transcripts.
  • Developed and maintained algorithmic trading bots that operated 24/7 on the Binance exchange, trading a wide range of cryptocurrencies. Utilized Python for development and Appsmith to display analytics of the trading bot.
Placeholder image

ResoluteAI Software

Machine Learning Engineer Intern — to

  • Built a multilingual OCR pipeline leveraging image preprocessing and EasyOCR to identify and extract package labels for Daawat, overcoming challenges related to variable video quality and lighting.
  • Developed a computer vision solution to accurately count defective towels on a moving conveyor, enhancing quality control processes in high-volume production environments.
  • Implemented an active learning pipeline that reduced data annotation time by 70%, optimizing model training for defect detection in large-scale manufacturing at Welspun.

Education

Dayananda Sagar College of Engineering

BE in Computer Science -

  • Ranked among the top 5% of my batch.
  • Worked as a Research Assistant under Prof. Swetha M. D., focusing on automating the diagnosis of MS lesions in normally appearing white matter in collaboration with doctors from NIMHANS .
  • Collaborated with Prof. Vindhya as a co-guide for final-year capstone projects in the AIML department.

Academic Service

Served as a reviewer for prestigious conferences, including ACM Multimedia 2023, ACML 2023, AI-ML Systems 2023, WACV 2024, CVPR 2024, ECCV 2024, ACML 2024, WACV 2025, and ICLR 2025.