We present a novel framework to track and catch reactive objects in a dynamic 3D world.

We present a novel framework to simultaneously learn the geometry, appearance, and physical velocity of 3D scenes.

We propose a novel ray-based 3D shape representation, achieving a 1000x faster speed in rendering.

We propose the first unsupervised 3D semantic segmentation method, learning from growing superpoints in point clouds.

We introduce a single pipeline to simultaneously reconstruct, decompose, manipulate and render complex 3D scenes.

We propose a generalizable framework for robotic manipulation.

We introduce the first unsupervised 3D object segmentation method on point clouds.

We systematically investigate the effectiveness of existing unsupervised models on challenging real-world images.

We introduce a simple weakly-supervised neural network to learn precise 3D semantics for large-scale point clouds.

The journal version of our SpinNet. More experiments and analysis are included.

We propose a new method to recover the geometry and semantics of continuous 3D scene surfaces from point clouds.

The journal version of our SensatUrban. More experiments and analysis are included.

We present a learning-based LiDAR relocalization framework to efficiently estimate 6-DoF poses from LiDAR point clouds.

We introduce a simple implicit neural function to represent complex 3D geometries purely from 2D images.

The journal version of our RandLA-Net. More experiments and analysis are included.

We introduce a simple and general neural network to register pieces of 3D point clouds.

We introduce an urban-scale photogrammetric point cloud dataset and extensively evaluate and analyze the state-of-the-art algorithms on the dataset.

We introduce a simple end-to-end neural network with self-attention to estimate global poses from FMCW radar scans.

We introduce an efficient and lightweight neural architecture to directly infer per-point semantics for large-scale point clouds.

We propose a simple and efficient neural architecture for accurate 3D instance segmentation on point clouds. It achieves the SOTA performance on ScanNet and S3DIS (June 2019).

We propose a novel end-to-end deep parallel neural network to estimate the 6-DOF poses using consecutive 3D point clouds.

We propose an attentive aggregation module together with a training algorithm for multi-view 3D object reconstruction.

We propose a simple embedding learning method that jointly optimises for an auto-encoding reconstruction task and for estimating the corresponding attribute labels.

We propose a novel neural architecture to reconstruct the complete 3D structure of a given object from a single arbitrary depth view using generative adversarial networks.

We present a neural framework to predict how a 3D object will deform under an applied force using intuitive physics modelling.

We propose an efficient and holistic pipeline to simultaneously learn the semantics and structure of a scene from a single depth image.

We present a novel generative adversarial network to predict body deformations under external forces from a single RGB-D image.