Dipankar Nandi

Developed a room-scale geometry capture pipeline using ARKit and RoomPlan, converting spatial depth data into structured point clouds and refined mesh representations Trained and deployed 3D Gaussian Splat models, integrating splat-based photorealistic rendering into a Three.js web viewer. Created a collision-aware navigation system using BVH-accelerated physics. Facilitated human-like room-to-room traversal. Designed a 3D interactive object analyzer using SAM3D and Ray-tracing interaction.

Developed NeRF and Gaussian Splatting pipelines for a large-scale multi-camera 3D scanning system. Built a 2D–3D camera - LED calibration & tracking pipeline with custom markers and homography-based correction; reduced calibration time (~2.5 hours to ~30 minutes) Applied SfM and multi-view geometry to refine camera poses and mitigate drift, scale & alignment errors; reduced RMSE (~4.8 to ~2.1) Investigated reconstruction of reflective objects using hybrid rasterization and ray-tracing approaches

Generated synthetic 3D datasets and reconstructed indoor environments (~500 scenes) using Unity3D and NeRF pipelines Developed simulation datasets for autonomous driving perception experiments (~50–70 outdoor scenes under varied conditions) using Unity3D Built perception pipelines for 3D object recognition, segmentation, and tracking using point clouds, meshes, and monocular 3D detection methods Supported dataset preparation and annotation workflows for human pose and object understanding tasks

Built a NeRF and Gaussian Splatting based large-scale dataset (~600K images) for fisheye-view human reconstruction and pose estimation Trained neural rendering and pose estimation models using multi-GPU workflows, across multiple poses without full retraining Integrated SAM, GroundingDINO, and Vision Transformers for automated human segmentation and scene understanding Generated SMPL-X / SMPL-H meshes and evaluated downstream 2D/3D pose estimation pipelines for reconstruction accuracy and robustness

Project1 Annotated a TopView Dataset using CVAT and further corrected it by 30% Implemented a Deep Learning pipeline with CenterNet for Human Pose and Object Detection Refined the accuracy from the initial 33% to around 51% and maximized to 65% Project2 Assisted in Soil moisture sensor data collection and processing using SQL and Pandas Sensor Data reporting with Microsoft Power BI Developed Random Forest, SVM, and CNN predictive models with an accuracy of around 67%, 56%, and 73% respectively

Assisted in Camera Calibration and Data correction on omnidirectional images, improving the dataset by 40% Presented a comparative study between SVM, K-means, and CNN algorithms Achieved a precision value of 60% for Human Activity Recognition by constructing a CNN-based model

Executed Data collection and cleaning for ANN and CNN architectures for Object Recognition and Localization Deployed a YOLOv3 architecture to show an improved accuracy of 20% over Fast-RCNN implementation

2 Projects: Parking Space Detection and Customer Sentiment Analysis Engineered a Parking Inspection system using Image processing(OpenCV) and Machine Learning algorithms Integrated the Parking system within an IoT environment(LoRaWAN) to ease the tracking process by 45% Improved market analysis accuracy by 25% through SQL, Pandas and Seaborn Visualized and prepared a dashboard using Power BI Developed 2 Regression-based models with an accuracy of 73% to improve customer-market sentiment

Thesis ( Ongoing ): Top-View Synthesis and Ground Truth Estimation using Neural Field Radiance on Humans CourseWork: Computer Vision and Neural Networks Programming and Data Analysis Image processing and Pattern Recognition