Saketh Bachu
Updates
Work Experience
Publications
Projects
I am a grad researcher at the University of California at Riverside (advised by Prof. Amit Roy Chowdhury). I previously worked as a Research Assistant (advised by Prof. Vineeth N Balasubramanian's) at the Indian Institute of Technology, Hyderabad (IITH). My fields of research broadly include Transfer Learning, Causality, ExplainableAI, and Learning under Limited Supervision.

Before that, I was associated with Mercedes Benz Research and Development, Bangalore, India where I worked on improving Lane Marking Detection in extreme weather conditions. I was advised by Dr. Deepak Kumar Panda and his team.

I completed my undergrad degree from Visvesvaraya National Institute of Technology, Nagpur, where I was affiliated with IvLabs, VNIT Nagpur (a student-run AI and Robotics club) supervised by Prof. Shital Chiddarwar.

My current interests broadly include the development of LLMs/VLMs and the deployment of Agentic AI solutions for real-world problems. I am currently on the job market. Please reach out if you are hiring for Machine Learning Engineers, Data Scientists, or Applied Scientists.

News
Jul '25
Excited to give a talk at Mayachitra Inc on "Advances in Safety Alignment of Vision Language Models (VLMs)".
Jul '25
I successfully defended my MS thesis titled "On the Risks of Generative AI" at UC Riverside!
Jun '25
Our paper "VOccl3D: A Video Benchmark Dataset for 3D Human Pose and Shape Estimation under Occlusions" is accepted at ICCV 25.
Mar '25
Our paper "Examining Safety Alignment Across Image Encoder Layers in VLMs" is accepted at ICML 25.
Jan '25
Our paper "Causal Order: The Key to Leveraging Imperfect Experts in Causal Inference" is accepted at ICLR 25.
Oct '24
Our paper "STRIDE: Single-video based Temporally Continuous Occlusion-Robust 3D Pose Estimation" is accepted at WACV 25.
Sep '24
I will be a Teaching Assistant for a course on "Introduction to Robotics and Artificial Intelligence" for Fall 24!
Jun '24
Excited to join JpMorgan Chase & Co. as an AI & DS Summer Associate 2024 (Consumer and Community Banking Risk Modelling Team).
Mar '24
Our paper "Active Transferability Estimation" is accepted at L3D-IVU Workshop at CVPR 24.
Jan '24
Our paper "Causal Inference using LLM-Guided Discovery" is accepted at LLM-CP Workshop at AAAI 24.
Dec '23
Our paper "Towards Learning and Explaining Indirect Causal Effects in Neural Networks" is accepted at AAAI 24.
Jul '23
Our paper "Building a Winning Team: Selecting Source Model Ensembles using a Submodular Transferability Estimation Approach" is accepted at ICCV 23.
Jul '23
Delighted to receive the 'Outstanding Reviewer Award' at MLRC 22.
Mar '23
Our paper "How Well Do Existing Partial Label Learning Methods Adapt to Class Incremental Learning Tasks?" is accepted at CLVISION 23 (CVPR 23).
Feb '23
Jan '23
Preprint of our new paper on "Towards Estimating Transferability using Hard Subsets" is now out!
Dec '22
I am excited to volunteer and attend ACML 2022. Excited to meet and connect with brilliant researchers.
July '22
I will be Taking the "Deep Learning for Computer Vision Course (DL4CV)" offered by IITH on NPTEL.
June '22
May '22
Elated to share that I will be joining "Prof. Vineeth N Balasubramanian's Lab at IIT Hyderabad" as a Research Assitant!
Mar '22
Nov '21
Excited to start an internship at "Mercedes-Benz Research and Development, India"! Will be working on vehicle vision.
Nov '21
Our paper "Go with the flow" is accepted at CSS Workshop at ICLR'22.
Sep '21
Started working as a guest researcher at Universität Osnabrück, Bio-Inspired Computer Vision lab.
Aug '21
Selected for Online Asian Machine Learning school at ACML 2021.
Jun '21
Elated to start a research internship (remote) at DFKI, TU Kaiserslautern!
Jun '21
Our results paper, "Paying Attention to Video Generation" is accepted to be included in the PMLR journal!
Dec '20
Our paper, "A review of video generation approaches" is accepted at the ICETEST 2020 conference.
Nov '20
Our proposal, "Paying Attention to Video Generation" is accepted at Preregistration Workshop at NeurIPS'20.
Jun '20
Won the second runner-up prize at the Reinvent - Corona hackathon.
May '20
Accepted an internship oppurtunity at Drona Aviation. (update: postponed due to COVID-19 outbreak)
May '20
Our paper on "Prediction of COVID-19 Outbreak" is accepted at the ICFE 2020 conference.
Sep '19
Check out the code for my internship project titled "Face Liveness Detection", which has 50+ stars on Github.
May '19
Starting an internship position at SimpleCRM, Nagpur, excited to work on AI powered Business tools.
Work Experience
Oct '23 - July '25
Vision and Learning Group, UC Riverside - Graduate Student Researcher.
June '24 - Aug '24
JPMorgan Chase & Co, USA - AI & DS Summer Associate.
May '22 - Sep '23
Nov '21 - May '22
Mercedes-Benz Research & Development, Bangalore - Deep Learning Research Intern.
Sep '21 - Dec '21
May '21 - Dec '21
DFKI, TU Kaiserslautern. - Deep Learning Intern.
May '19 - May '22
IvLabs, VNIT Nagpur - Core Comittee Member.
May '19 - Oct '19
SimpleCRM, Nagpur - Deep Learning Intern.
Publications
Layer-wise Alignment: Examining Safety Alignment Across Image Encoder Layers in VLMs
Collaborators: Erfan Shayegani, Rohit Lal, Trishna Chakraborty, Arindam Dutta, Chengyu Song, Yue Dong, Nael Abu-Ghazaleh, Amit K. Roy-Chowdhury

TLDR: We identify a novel safety vulnerability in Vision-Language Models that arises when exiting early from the image encoder. We further show that this vulnerability can be mitigated through a simple modification to the PPO algorithm used in RLHF. This work was presented as a spotlight poster at ICML 2025.

Venue: International Conference of Machine Learning (ICML) 2025 - Spotlight 🏆
VOccl3D: A Video Benchmark Dataset for 3D Human Pose and Shape Estimation under real Occlusions
Collaborators: Yash Garg, Saketh Bachu, Arindam Dutta, Rohit Lal, Sarosij Bose, Calvin-Khang Ta, M. Salman Asif, Amit Roy-Chowdhury

TLDR: We propose VOccl3D, a large-scale synthetic video dataset specifically designed for training and evaluating algorithms for 3D human pose and shape estimation (HPS) in realistic occlusion scenarios.

Venue: International Conference on Computer Vision (ICCV) 2025
STRIDE: Single-video based Temporally Continuous Occlusion-Robust 3D Pose Estimation
Collaborators: Rohit Lal, Yash Garg, Arindam Dutta, Calvin-Khang Ta, Hannah D. Cruz, Dripta S. Raychaudhuri, M. Salman Asif, Amit K. Roy-Chowdhury

TLDR: we propose STRIDE (Single-video based Temporally continuous Occlusion-Robust 3D Pose Estimation), based on a novel Test-Time Training (TTT) approach for estimating 3D human poses under significant occlusions.

Venue: Winter Conference on Applications of Computer Vision (WACV) 2025
Building a Winning Team: Selecting Source Model Ensembles using a Submodular Transferability Estimation Approach
Collaborators: Vimal K B, Tanmay Garg, Niveditha Lakshmi Narasimhan, Raghavan Konuru, and Vineeth N Balasubramanian

TLDR: We propose a novel transferability estimation metric dubbed "OSBORN" for measuring the transfer learning performance of source model ensembles without actually fine-tuning them.

Venue: International Conference on Computer Vision (ICCV) 2023
Multi-stage Attention-Pooling Network for Lane Marking Detection
Collaborators: Tushar Garg, Deepak Panda, Mallikarjuna Reddy, Shaikh Ibrahim, Bharath Bhat

TLDR: An attention-based multi-stage neural network built to reduce the false-positive rate and produce compact lane marking segmentations.
Multi-Stage Pyramid Parsing Network For Lane Marking Detection
Collaborators: Tushar Garg, Deepak Panda, Mallikarjuna Reddy, Shaikh Ibrahim, Bharath Bhat

TLDR: A combination of spatial pyramid pooling and gated attention bocks are used to produce semantically consistent lane marking segmentations.
Go with the Flow: the distribution of information processing in multi-path networks
Collaborators: Mats Leon Richter, Krupal Shah, Anna Wiedenroth, Ulf Krumnack

TLDR: Analyzing the information processed in different parallel pathways of neural networks.
Paying Attention to Video Generation
Collaborators: Rishika Bhagwatkar, Khurshed Fitter, Akshay Kulkarni, Shital Chiddarwar

TLDR: Generating new video frames using GPT and a novel Autoencoder.
Selected Projects
Covid19 Outbreak and Mitigation Prediction
TLDR: An AI pipeline designed to forecast the number of covid19 cases and also predicts suitable mitigation measures to control the outbreak.
[Code]
Face Liveness Detection
TLDR: A neural network titled ‘CRMNet’ that is designed to tackle face spoofing attacks.
[Code]
https://vnit.ac.in/