Yi Wei

I am a research engineer in Huawei, working on computer vision, including 3D reconstruction, video generation and world model. I obtained my Ph.D degree at the Intelligent Vision Group (IVG), Department of Automation, Tsinghua University, advised by Prof. Jiwen Lu.

Prior to that, I received my Bachelor's degree from the department of Electronic Engineering, Tsinghua University in 2019 (Ranking 6/245). I have also spent some time at DeePhi Tech (Xilinx), Sensetime , Microsoft Research Asia, XPeng, ByteDance, PhiGent Robtics, Gaussian Robotics and Apple.

Email  /  Google Scholar  /  Github  /  Twitter  /  Curriculum Vitae

* indicates equal contribution

We introduce the Geometry-Aware Large Reconstruction Model (GeoLRM), an approach which can predict high-quality assets with 512k Gaussians and 21 input images in only 11 GB GPU memory.

We propose an OccNeRF method for self-supervised multi-camera occupancy prediction, which adopts the parameterized occupancy fields, multi-frame photometric loss and open-vocabulary 2D segmentation.

We propose Sherpa3D, a new text-to-3D framework that achieves high-fidelity, generalizability, and geometric consistency simultaneously.

We propose a SurroundOcc method to predict the volumetric occupancy with multi-camera images and generate dense occupancy ground truth with sparse LiDAR points.

Towards a comprehensive benchmarking of surrounding perception algorithms, we propose OpenOccupancy, which is the first surrounding semantic occupancy perception benchmark.

We propose Deformable PV-RAFT, where the Spatial Deformation deforms the voxelized neighborhood, and the Temporal Deformation controls the iterative update process.

Beyond NerfingMVS, we further present NerfingMVS++, where a coarse-to-fine depth priors training strategy is proposed to directly utilize sparse SfM points and the uniform sampling is replaced by Gaussian sampling to boost the performance.

We propose the LiDAR Distillation to bridge the domain gap induced by different LiDAR beams for 3D object detection.

We propose a SurroundDepth method to incorporate the information from multiple surrounding views to predict scale-aware depth maps across cameras.

We present a new multi-view depth estimation method that utilizes both conventional SfM reconstruction and learning-based priors over the recently proposed neural radiance fields (NeRF).

We propose a novel solver that iteratively solves for per-view depth map and normal map by optimizing an energy potential based on the locally planar assumption.

We present point-voxel correlation fields for 3D scene flow estimation which migrates the high performance of RAFT and provides a solution to build structured all-pairs correlation fields for unstructured point clouds.

We propose a weakly supervised 3D detection method without using 3D labels, which consists of coarse 3D segmentation and 3D bounding box estimation two stages.

we present a new perspective towards image-based shape generation. Unlike most single-view methods which are sometimes insufficient to determine a single groundtruth shape because the back part is occluded, our method levergae multi-view consistency for 3D reconstruction.

we propose a simple and general framework for training very tiny CNNs for object detection. Our method leverages the fact that mimic and quantization can facilitate each other.

We propose the Two-Stream Binocular Network (TSBnet) to detect fingertips from binocular images. Different with previous depth-based methods, we directly regress 3D positions of fingertip from left and right images.

Website Template