Interactive 4D (Dynamic) Point Cloud Retrieval System

July 2026

#4D Vision #CL4D #Dynamic Point Clouds #Vector Database #Python #AI #Sasika Amarasinghe

A vector database and search system for retrieving 4D dynamic point clouds using natural language and multimodal queries, developed by Sasika Amarasinghe as part of the CL4D research project at 4D Vision UoM.

This project implements an Interactive 4D (Dynamic) Point Cloud Retrieval System designed to index and search spatio-temporal 3D representations (dynamic point clouds over time) using natural language commands.

The system leverages the CL4D (Contrastive Language–4D Pretraining) model to encode dynamic point clouds and text queries into a shared embedding space, allowing users to query complex 3D human actions and physical interactions.

🌐 Official Project Website & Research: Explore the complete research paper, benchmarks, and dataset on the official project page:
👉 4D Vision UoM — CL4D: Contrastive Language–4D Pretraining for Vision-Language Reasoning in Dynamic Scenes

📐 System Architecture

Interactive 4D Point Cloud Retrieval System Overview

System architecture overview: Spatio-temporal sequences of action point clouds are encoded using the CL4D Encoder, stored in a Vector Database, and queried interactively via natural language commands mapped to text embeddings. Developed by Sasika Amarasinghe at University of Moratuwa.

🎬 Demonstration Video

Below is the demonstration of the interactive query processing and retrieval visualization interface:

Interactive query retrieval visualization showing dynamic point cloud frames (e.g., drilling a hole) matched to a user command.