email: aknihdkv@wendra.ucm.ude
I am a Master's student in Computer Vision (MSCV) at the School of Computer Science (Robotics Institute), Carnegie Mellon University, where I work with Prof. Andrej Risteski on the theory of generative modeling and sampling, and with Prof. László A. Jeni on object-centric 4D reconstruction. I am also a Machine Learning Engineering Intern at Sync Labs, working on video generative AI and lip-sync pipelines. Earlier at CMU, I worked with Prof. Fernando De la Torre on diagnosing aerial object detectors with foundation generative models.
Before CMU, I spent several years at IIIT Hyderabad. As a Research Scientist at CVIT, I worked with Prof. C. V. Jawahar, Prof. Vineeth N. Balasubramanian, and Prof. Chetan Arora on autonomous-driving perception, pedestrian-intent prediction, and domain-adaptive detection, work that led to publications at IROS 2024 and WACV 2022, one U.S. patent, and two Indian patents. As an Applied AI Researcher at Product Labs, I also contributed to Bhashini, the Government of India's national language-AI mission. I hold a B.E. in Computer Science from Savitribai Phule Pune University.
Outside research, you'll usually find me at the piano, singing, swimming, running, meditating, or lost in a good book, with a soft spot for भक्ती संगीत, अभंग, and sci-fi (Iron Man, always). Giving back matters to me too: I've mentored students through AI4ALL and taught schoolkids to code.
My path into research started with a question I couldn't let go of as an undergrad: what would the next revolution in computing really look like? I was fascinated by the mathematics beneath how things work, especially the structure and uncertainty behind intelligent behavior. Half-seriously and half inspired by Iron Man's Jarvis, I became captivated by the idea of mind-reading computers: machines that could grasp intent and understand the why behind what people do. That, to me, was the frontier worth chasing.
That fascination with intent grew into a research program. At CVIT, I worked on reading pedestrian intent and modeling object-scene interaction, using causal, explainable models to predict not just what an agent will do, but why, alongside earlier work on domain adaptation, spatio-temporal reasoning with graph networks, and generative augmentation, all tied to the broader question of how machines can understand behavior in complex, changing environments.
Over time, this question pulled me one level deeper. To understand why agents act, we also need to understand the worlds they act within: the geometry of a scene, how objects move, how interactions unfold, and how these dynamics persist over time. This shift drew me toward computer vision, generative modeling, and spatial intelligence, especially models that go beyond simply perceiving a scene, and instead reconstruct how the world unfolds across space and time.
Right now I'm pursuing two complementary threads: with Prof. Andrej Risteski, the theory of reward-tilted sampling, on how to steer generative models toward desired outcomes in a principled way; and with Prof. László A. Jeni, Slot4D, an object-centric approach to 4D reconstruction that links low-level geometric tracking to object-level scene understanding.
Across these projects, I am interested in a broader question: how can we build models that move beyond perception alone: models that understand structure, dynamics, intent, and the physical worlds in which intelligent behavior unfolds?
Google Scholar · Semantic Scholar · ORCID
Diagnosing Aerial-View Object Detectors with Foundational Image Generative Models
ECCV 2026European Conference on Computer Vision
A synthetic-to-real diagnostic framework that uses foundation image generative models to build attribute-controlled aerial testbeds, exposing systematic detector weaknesses across scene type, season, and weather, and guiding targeted data collection for up to 11% AP50 improvement.
Can Reasons Help Improve Pedestrian Intent Estimation? A Cross-Modal Approach
IROS 2024IEEE/RSJ International Conference on Intelligent Robots and Systems
Introduced MINDREAD, a cross-modal multi-task framework that predicts not just what a pedestrian will do but why, alongside the reason-enriched PIE++ dataset (~909K textual explanations), yielding +5.6% accuracy and +7% F1 over baselines.
To Miss-Attend Is to Misalign! Residual Self-Attentive Feature Alignment for Adapting Object Detectors
WACV 2022IEEE/CVF Winter Conference on Applications of Computer Vision
ILLUME, a residual self-attentive feature alignment method whose Self-Attention Feature Map (SAFM) module attends to object-related regions, producing domain-invariant features and state-of-the-art adaptive object detection.
GAMMA: Generative Augmentation for Attentive Marine Debris Detection
arXiv 2022
An efficient generative augmentation approach that uses CycleGAN to translate abundant in-air plastic imagery into underwater-style images, addressing the scarcity of underwater debris data for visual detection.
Building video-generative-AI and lip-sync systems, focused on vision-language models (VLMs) for automated visual quality control, replacing manual, frame-by-frame review of AI-generated video. Architected GPU-served VLM inference on Modal: deployed multimodal models (Google DeepMind's Gemma-3-12B, Qwen2.5-VL-7B), benchmarked against NVIDIA Cosmos Reason, and engineered structured-output prompting, multi-model agreement gating, and per-model routing around each model's blind spots. Built a human-in-the-loop labeling loop feeding a LoRA fine-tuning path for Qwen2.5-VL on artifact crops, and evaluated RLHF vs. supervised fine-tuning with a per-class precision/recall/F1 harness to drive model iteration.
Studying reward-tilted sampling: principled methods for steering generative models toward target distributions, with a focus on stable, structured sampling.
Developing Slot4D, an object-centric approach to 4D reconstruction: built an adaptive slot-attention head over frozen geometric backbones (VGGT / VGGT-Omega) that decomposes scenes into object slots and masks, plus a slot-conditioned refinement network that has delivered up to 4× better surface reconstruction on CO3D, ScanNet, and Virtual KITTI; now extending these gains to multi-view-consistent, streaming 4D.
Used foundation generative models to diagnose and improve aerial-view object detectors under controlled visual attributes, designing synthetic data workflows to expose systematic detector failures across scale, viewpoint, background, occlusion, and appearance.
Containerized OCR models with Docker and deployed via NVIDIA Triton, achieving real-time latency (0.11s printed, 0.051s handwritten) across 22 Indian languages.
Where my research really began: leading work on pedestrian intent prediction (MINDREAD, IROS 2024) and domain-adaptive object detection (ILLUME, WACV 2022), building large-scale CV pipelines for detection, intent estimation, and multimodal learning.
Earlier, I spent two years as a Software Developer at Barclays (Pune) before fully turning toward research.
Slot4D explores object-centric learning for 4D scene reconstruction, built on feed-forward geometric foundation models (VGGT / VGGT-Omega). The first stage of the project built an adaptive slot-attention head that groups the backbone's features, grounded in predicted depth and camera geometry, into per-object slots and masks, reliably decomposing scenes into objects and regions. Building on this, a slot-conditioned refinement network predicts per-object corrections to the base geometry and has delivered up to 4× better surface reconstruction on CO3D, ScanNet, and Virtual KITTI, even closing holes the base model leaves behind. The current focus is carrying these per-view gains over to a globally consistent multi-view reconstruction, en route to the project's larger goal: dynamic, streaming 4D reconstruction that sees objects, not just points.
Studying how generative models can be steered toward high-reward outputs by reshaping their sampling distribution. This project investigates reward-tilted distributions, p*(x) ∝ p(x) exp(r(x)), and lift-and-tilt constructions for making reward-guided generation more stable and principled, especially when the reward landscape is complex, indefinite, or non-concave.
This project tackles 3D Gaussian Splatting when the input views come at very different resolutions, for example a few sharp, high-resolution captures alongside many low-resolution images of the same scene. It studies how to fuse these heterogeneous sources into a single consistent radiance field, using resolution-aware, detail-preserving densification so that fine detail from the high-resolution views is retained rather than washed out by the coarser ones. The goal is high-fidelity reconstruction that makes the most of mixed-quality, real-world image collections.
Estimating the full pose of a robot, its joint configuration and 2D/3D keypoints, from a single RGB image without depth or known camera-to-robot calibration. The project builds a holistic pipeline trained largely on synthetic data, using domain randomization and synthetic-to-real transfer to bridge the appearance gap and generalize across camera viewpoints and even to robot types unseen during training. Robust single-view pose estimation like this is a step toward lighter-weight perception for manipulation and human-robot interaction.
Classification of a remote-sensing image dataset using a fusion model that combines spatial CNN features with DCT features. A three-layer CNN fusion model with DCT and LBP improves prediction accuracy on the UC Merced LandUse dataset.
Aspect-based sentiment analysis on Net Promoter Score (NPS) survey data for a retail online-banking platform, to understand customer reviews of features like Payments or Homepage. Built an AngularJS dashboard visualizing per-feature sentiment, helping identify improvements by analyzing negative reviews for specific features.
A user-friendly techno-quiz application, built from scratch in Turbo C++ using data structures and OOP, that tests your knowledge across three levels: Computer World Tycoons, Programming Basics, and "Bugzzz… Find me if you can!", advancing only after you earn a badge for the current level.
An Android application to track loans recollected from customers, with two companion apps for owner and employee roles. The owner manages employees and transaction history, while employees are allocated daily collection accounts reflected in a realtime database, cutting down on paperwork.
An Android app that helps people connect with NGOs by easily finding nearby organizations and their information, so they can support local causes through donations or activities like birthday celebrations, contributing toward better living in their community.
A movie-ratings analysis tool built with Qt Creator and MySQL, where a user's ratings and comments for a film of a given year dynamically update the movie's overall rating.
I occasionally write to share what I've learned, from intuitions behind machine-learning concepts to helping others navigate grad-school applications and find their own path.
Building support vector machines up from basic geometry to how the algorithm finds the boundary that maximizes the margin between classes.
Read on Substack → SubstackAn intuitive walk through the math that finally made logistic regression click, from odds and log-odds to the sigmoid and why it works.
Read on Substack → MediumMy full Statement of Purpose, application tips, and honest reflections on applying to competitive computer-vision, AI, and robotics programs.
Read on Medium → MediumA step-by-step guide to identifying your research interests, connecting with faculty, and building a focused, coherent application.
Read on Medium → MediumThe tools I used, section-by-section practice strategy, and the small tricks that made a big difference in getting a strong score.
Read on Medium →