- Implemented unsupervised LiDAR 3D object detection and scene understanding with DBSCAN and K-means + PCA, improving perception in unstructured environments.
- Reimplementing state-of-the-art BEVFusion and self-supervised learning methods as a foundation for advancing camera–LiDAR fusion.
About
I'm Bipin Saha (please call me Bipin). I grew up in Bangladesh and now live in Greensboro, North Carolina. My work sits at the intersection of computer vision and robotics, and has moved from deploying real-world AI systems as a Machine Learning Engineer to researching the perception stack that autonomous vehicles depend on.
As a PhD student at the ACCESS Lab, North Carolina A&T State University, I focus on visual navigation and camera–LiDAR fusion for autonomous systems, with the goal of making self-driving perception more perceptive, robust, and explainable. I am particularly interested in self-supervised perception, multimodal sensor fusion, and reasoning-driven robotic decision-making.
Before the PhD, I worked on the Industry 4.0 team at Business Automation Limited, building and deploying systems across computer vision, vision–language models, LLMs, and vision-enabled robots. That experience shapes how I approach research — I care about methods that are scalable, explainable, and deployable. Long term, I am aiming for a research career in autonomous-driving perception, in industry or academia.
Bipin (বিপিন) — from Sanskrit বিপিন / विपिन, “a vast natural grove.”
Research focus
- Multimodal fusion Camera–LiDAR fusion for 3D detection and scene understanding — reimplementing and extending BEVFusion and related methods at the ACCESS Lab.
- Data-efficient perception Self-supervised and unsupervised methods — LiDAR clustering, representation pretraining — that reduce dependence on labeled data for perception in unstructured environments.
- Robust & explainable detection Vehicle detection under adverse weather, complex scenes, and underrepresented regions, with explainable-AI diagnostics and benchmark datasets (BNVD).
- ACCEPTED IEEE QPAIN 2026 paper benchmarking CNN, Transformer, and VLM detectors for complex-scene vehicle detection.
- 52.68% reduction in manual workload from a handwritten-prescription OCR pipeline deployed across 64k+ patients.
- ~98% accuracy in appearance-based eye-gaze estimation across 13 directional targets (IC4ME2 2021).
News
IEEE QPAIN 2026 acceptance New
“Comparative Evaluation of CNN, Transformer, and Vision-Language Approaches for Vehicle Detection in Complex Scenarios.”
Joined the ACCESS Lab as a Graduate Research Assistant
Started the PhD in Electrical and Computer Engineering at North Carolina A&T State University.
Publications
Selected work below — full list on Google Scholar. Bipin Saha in bold.
Selected
IEEE QPAIN 2026 · accepted Feb 2026
A practical benchmark framework for complex-scene vehicle detection across CNN, Transformer, and vision-language model families.
IEEE Int. Conf. on Quantum Photonics, Artificial Intelligence (QPAIN) · Jul–Aug 2025
Conference paperIEEE Xplore: 29 Sep 2025DOI
Robust adverse-weather detection with explainability, aimed at trust in safety-critical deployments.
Preprint · benchmark datasetarXiv: 20 May 2024DOI
A region-specific benchmark dataset to improve generalization and fairness in vehicle detection.
2021 Int. Conf. on Computer, Communication, Chemical, Materials and Electronic Engineering (IC4ME2)
Conference paperIEEE Xplore: 10 May 2022DOI
Additional
PreprintDOI
Discover Analytics 3 (1), 9 · 2025
Journal paperDOI
2024 Int. Conf. on Recent Progresses in Science, Engineering and Technology
Conference paperDOI
AIP Advances 13 (2) · 2023
Journal paperDOI
Experience
Research
Industry
- Digitized handwritten prescriptions with layout segmentation and OCR (CER 0.0951), fine-tuned LLMs for structured output, cut inference to ~2.28 s, and reduced human workload by 52.68% — impacting 64k+ patients.
- Built a Bangla law-consultancy chatbot on Llama-3 with an agentic RAG system and TF-IDF indexing; reduced training memory 3.2× via QLoRA / QDoRA / PEFT; 20% faster inference.
- Developed a multi-agent code-suggestion framework using LangGraph state memory, DeepSeek-R1, and a Milvus index built from GitLab issue text.
- Led an object-detection system using YOLOv5 / YOLOv8 for 20+ grocery products, reaching mAP50 0.955 and mAP50:95 0.872.
- Deployed production-ready detection and recognition models for diverse retail items.
- Designed a 2R inverse-kinematics robotic arm and tuned PID control for precise slot navigation in a 2D plane.
Undergraduate Research
- Designed and simulated an n-ZnSe / p-CuSbSe2 / p+-CGS device in SCAPS-1D.
- Studied absorber thickness, doping, and defect density; reached 43.77% simulated efficiency (vs. 27.74% single-heterojunction).
- Published in AIP Advances 13, 025255 (2023).
- Built a real-time eye-gaze interface using active appearance methods with commodity cameras, accounting for natural head motion.
- Separated sclera pixel areas from masked eye images; evaluated decision-tree, random-forest, and extra-tree classifiers.
- Reached ~98% accuracy across 13 gaze targets; published at IC4ME2 2021.
Skills
-
Perception & robotics
3D object detection (LiDAR) · camera–LiDAR / BEV fusion · self-supervised & unsupervised learning · visual navigation · path planning (RRT) · inverse kinematics · PID control · imitation learning
-
Vision & ML
PyTorch · TensorFlow / Keras · scikit-learn · OpenCV · object detection (YOLOv5/v8) · OCR (PaddleOCR, TrOCR) · explainable AI
-
LLM / VLM systems
Fine-tuning (QLoRA, QDoRA, PEFT) · agentic & semantic RAG · LangGraph · Milvus · Llama-3, Mistral, Gemma, DeepSeek-R1
-
Datasets & tooling
Dataset construction & curation (BNVD, FORS-EMG) · reproducible benchmarking · CoppeliaSim · SCAPS-1D · Arduino / embedded C++ · multithreaded real-time pipelines
-
Languages
Python · C++ · Lua
Projects
-
End-to-End Behavioral Cloning for Self-Driving (Udacity)
Imitation learning · TensorFlow · PID control
A CNN maps camera frames directly to steering commands while a PID controller regulates speed; gain tuning yields smooth, stable driving in simulation.
Stable lane-following and smoother steering after PID gain optimization.
-
RRT Path Planning from Scratch
NumPy · PyGame
A full Python implementation of the Rapidly-Exploring Random Tree algorithm for 2D planning, visualizing tree expansion and obstacle-aware navigation.
Collision-aware path construction with interpretable search behavior.
-
Slab Bend Detection & Tracking (with Tata Steel)
YOLO · Tkinter · multithreading
Real-time industrial monitoring across multi-camera feeds: YOLO inference flags slab-bending anomalies while a GUI logs camera ID, position, timestamps, and confidence with database persistence.
Reduced operator reaction time via real-time alerts and synchronized multi-camera evidence logging.
-
KUKA Robot Object Sorting on a Conveyor
CoppeliaSim · Lua · inverse kinematics
A KUKA manipulator sorts objects over a conveyor: color sensors detect items, a gripper picks them, and IK motion planning places them at designated locations.
Consistent pick-and-place via deterministic motion planning and sensor-triggered logic.
-
Handwritten Prescription Digitization
PaddleOCR · TrOCR · YOLO · Gemma-7B
Line-by-line layout analysis and OCR extract patient IDs, medication, dosage, frequency, and history from handwritten prescriptions; a generative model converts the text into structured clinical records.
CER 0.0951, inference ~2.28 s, 52.68% less manual workload across 64k+ patients.
-
Appearance-Based Eye Gaze Estimation & Multilingual Keyboard
scikit-learn · Dlib · Random Forest
Real-time gaze tracking with standard cameras: active appearance modeling and ensemble classifiers drive a multilingual on-screen keyboard.
~98% gaze-direction classification accuracy on a low-cost camera setup.
Also: ChatPDF research assistant with semantic RAG (Gemini-Pro, Llama-3) · text-to-text summarization with Pegasus / Mistral · Bengali sentiment analysis with CNN-BiLSTM (90.31% accuracy) · landmine-detection and disposal robot (Arduino, C++).
Technical Writeups
- From 2D Perception to BEV Fusion: An Implementation Roadmap
- Camera–LiDAR Fusion in Practice: Failure Modes and Design Patterns
- Self-Supervised Perception for Autonomous Driving: A Practical Overview
- Vision–Language Approaches for Vehicle Detection in Complex Scenarios
- How to Build Reproducible Perception Benchmarks for AV Research
Education
Focus: visual navigation and camera–LiDAR fusion for autonomous systems.
Undergraduate research on appearance-based eye-gaze classification and vehicle detection in underrepresented regions. Relevant coursework: Control Systems, Digital Signal Processing, Microprocessors & Embedded Systems, Power Electronics, Computer Networks.
Awards & Service
Awards
Service & leadership
Founded and led the IEEE Robotics and Automation Society Student Branch Chapter, running research and outreach activities in robotics and automation.
Secretary and Publicity Coordinator for branch events and international seminars.
Also: IEEE Brand Ambassador (2020–21); Director of Technology & Resource Development, Bangladesh Innovative Education Service (2018–19); science writing for Bigganbortika (2018–19).