Hardik Prajapati
PhD Student at UC Santa Barbara | Computer Vision & ML Researcher | GCP Certified
I am a Ph.D. student in Electrical and Computer Engineering at the University of California, Santa Barbara, specializing in Computer Vision and Machine Learning. I am advised by Prof. B.S. Manjunath at Vision Research Lab . My research is driven by the goal of developing sophisticated world models that enable autonomous systems to perceive, reason about, and predict the dynamics of complex visual environments.
My current work focuses on bridging the gap between high-level human intent and dense, structured scene understanding. By integrating interactive foundation models with relational reasoning, I build frameworks that transform sparse visual cues into pixel-accurate world-state representations. I am particularly interested in:
- Spatiotemporal Grounding: Developing systems like Click2Graph that allow users to guide scene graph generation through direct visual prompting.
- Consistent World Dynamics: Designing architectures that maintain temporal coherence in dynamic environments, ensuring a non-fragmented understanding of how entities and their interactions evolve over time.
- Predictive Generative Modeling: Advancing long-range video generation through explicit future-frame prediction to simulate and anticipate complex scene evolution.
Prior to my doctoral studies, I earned my Master of Science from the University of Southern California and gained industry experience at Mayachitra Inc. and Analytos. There, I built scalable ML pipelines and multi-modal fusion systems for real-world applications. I am passionate about pushing the boundaries of spatial intelligence and temporal consistency to create the next generation of predictive AI.
Interests
- Video Content Understanding
- Video Generation
- Multi-modal sensor fusion and alignment
- Interactive/Promptable computer vision systems
Education
- PhD in Electrical & Computer Engineering (2024 - Present)
University of California, Santa Barbara (UCSB), USA - MS in Electrical & Computer Engineering (2021 - 2022)
University of Southern California (USC), USA
GPA: 4.0/4.0 - B.Tech in Instrumentation & Control (2014 - 2018)
Nirma University, India
GPA: 8.34/10.0
| August 2025 | Joined Mayachitra Inc. as Machine Learning Software Intern. |
| Sept 2024 | Started PhD at UC Santa Barbara under advisement of Prof. B.S. Manjunath. |
| Jan 2024 | Joined Analytos as Junior AI Engineer. |
| May 2023 | Tiny ML Model paper published in APSIPA Transactions. |
| May 2023 | Awarded Outstanding Academic Achievement Award (top 2 from 200+ students in the ECE department). |
| Jan 2023 | S3I-PointHop paper accepted at IEEE ICASSP 2023. |
| Dec 2022 | Graduated from USC with 4.0 GPA. |
| May 2022 | Started research internship at USC Media Communications Lab under Prof. C.-C. Jay Kuo. |
| Jan 2022 | Started as Course Mentor for Machine Learning course at USC (110 students). |
| Jan 2021 | Started Master's at University of Southern California. |
2025
Click2Graph: Interactive Panoptic Video Scene Graphs from a Single Click
arXiv, 2025
Click2Graph is the first interactive framework for Panoptic Video Scene Graph Generation (PVSG) that unifies visual prompting with spatial, temporal, and semantic understanding. From a single user cue, such as a click or bounding box, Click2Graph segments and tracks the subject across time, autonomously discovers interacting objects, and predicts ⟨subject, object, predicate⟩ triplets to form a temporally consistent scene graph
TCDSG: An End-to-End Approach for Action Tracklet Generation
arXiv, 2025
TCDSG is a unified end-to-end framework that integrates detection, tracking, and interaction prediction across video sequences. TCDSG introduces two key innovations: a sequence-level bipartite matching strategy that enforces stable query assignments across frames to reduce tracklet fragmentation without post-processing, and temporally conditioned decoder queries that inject inter-frame feedback directly into decoding for improved stability and accuracy. Together, these mechanisms yield tR@50 39.1% on Action Genome.
2023
S3I-PointHop: SO(3)-Invariant PointHop for 3D Point Cloud Classification
IEEE International Conference on Acoustics, Speech, and Signal Processing (ICASSP), 2023
A rotation-invariant approach for 3D point cloud classification that assigns class labels to point cloud scans expressed in arbitrary coordinate systems. The method derives invariant representations by leveraging principal components, rotation invariant local/global features, and point-based eigen features, enabling robust classification regardless of coordinate system orientation.
Semantic Segmentation - KITTI Dataset
Implemented state-of-the-art semantic segmentation for autonomous driving scenes using DeepLab architecture. Achieved real-time performance on KITTI benchmark with optimized inference pipeline.
PyTorch • DeepLab • Computer Vision
3D Point Cloud Classification
Developed lightweight models for efficient 3D point cloud object recognition on ModelNet40. Focus on edge deployment with minimal computational requirements.
PyTorch • Point Cloud • Edge AI
Object Detection - PascalVOC
Comparative study of object detection architectures including YOLO and Faster R-CNN. Detailed ablation studies on backbone architectures and optimization strategies.
PyTorch • YOLO • Faster R-CNN
Image Classification - STL10
Deep learning-based image classification on STL10 dataset with various CNN architectures. Explored data augmentation and transfer learning techniques.
TensorFlow • CNNs • Transfer Learning
Auto-Wrapper Changeover System
Industrial automation project at Unilever reducing changeover time by 40%. PLC-based system with real-time monitoring and predictive maintenance.
PLC • Industrial Automation • IoT
Sales Forecasting Pipeline
End-to-end ML pipeline for sales forecasting improving decision-making by 20%. Deployed on GCP using VertexAI with automated retraining.
GCP • VertexAI • Time Series
| 2025 | Machine Learning Software Intern, Mayachitra Architected a world model integrating multi-modal data fusion (vision, text, and telemetry) for predictive situational awareness and real-time semantic retrieval. |
| 2024 - Present | PhD Candidate, UC Santa Barbara Research on video scene graphs and trajectory anomaly detection under Prof. B.S. Manjunath |
| 2024 | Junior AI Engineer, Analytos Developed sales forecasting and vehicle routing optimization systems |
| 2022 | Graduate Researcher, USC Media Communications Lab Designed lightweight 3D point cloud classification models under Prof. C.-C. Jay Kuo |
| 2021 - 2022 | Course Mentor - Machine Learning, University of Southern California Coached 110 students, designed homework and exam problems |
| 2019 - 2020 | Engineer, G6 SuperHomes Designed smart home automation prototype with voice and mobile control |
| 2018 - 2019 | Manufacturing Engineer, Unilever India Supervised 43 employees, improved efficiency through automation |