Sparse Coding of Dynamic Point Clouds for Scene Analysis and Reconstruction (SPYDER)

01 January 2022 → 31 December 2025
Research Foundation - Flanders (FWO)
Research disciplines
  • Engineering and technology
    • Computer vision
    • Image and language processing
    • Pattern recognition and neural networks
    • Data visualisation and imaging
machine learning point clouds computer vision deep learning
Project description

With recent expansion of sensors that capture depth information along with other visual cues, three-dimensional (3D) point clouds start to play a pivotal role in many applications, like autonomous navigation, robotics, virtual and augmented reality. A proliferation of devices such as RGB-D cameras and LiDAR leads to huge amounts of 3D data that is being captured and analysed by various machine vision systems. Since storing and processing 3D data points in their raw form quickly becomes a bottleneck of a processing system, designing compact, i.e., sparse representations to enable efficient storage and analysis on the fly, is a major challenge. Applying emerging deep learning models on point cloud data directly is not possible because they are not structured and not ordered. Especially challenging is processing of sequences of point clouds, also called dynamic point clouds. While many recognition tasks benefit from using temporal sequences of the monitored scene, processing of dynamic point clouds is very difficult because they are not represented on a regular spatio-temporal grid. Another crucial problem is how to analyse raw point clouds without having to rely on many annotated examples. There is a high demand for such generative models for point clouds. The SPYDER project aims to solve these challenges by developing a generic framework for generative and scalable models for static and dynamic point clouds.