Shaohui Liu

ParticleSfM: Exploiting Dense Point Trajectories for Localizing Moving Cameras

, Shaohui Liu, , , ,

ECCV, 2022.         Part of my intern project at .

[Project]  [arXiv]  [YouTube]  [Slides]  [Code]  [BibTex] 

Connecting, optimizing dense point trajectories from pairwise optical flows and segmenting them into labeled point tracks benefit an effective global structure-from-motion system for dynamic in-the-wild videos (image sequences) that exhibit complex foreground motions.

A Confidence-based Iterative Solver of Depths and Surface Normals

*, Shaohui Liu*, , , ,

ICCV, 2021.         Part of my intern project at .

[Project]  [arXiv]  [Supplement]  [YouTube]  [Slides]  [Code]  [BibTex] 

A new iterative solver built upon locally planar assumption for confidence-based joint filtering of depths and surface normals exhibits both effectiveness and differentiability, which can not only serve as a post-processing step but also be integrated into deep multi-view stereo pipelines.

NerfingMVS: Guided Optimization of NeRF for Indoor Multi-view Stereo

, Shaohui Liu, , , ,

ICCV, 2021.         Accepted as an Oral.

[Project]  [arXiv]  [Supplement]  [YouTube]  [Code]  [BibTex] 

Integrating conventional SfM + MVS reconstruction and learning-based depth priors largely improves the learned geometry of neural radiance fields (NeRF) in general indoor scenes, leading to significantly better multi-view depth estimation and novel view synthesis.

DIST: Rendering with Differentiable Sphere Tracing

Shaohui Liu, , , , ,

CVPR, 2020.         Most work was done at .

[Project]  [arXiv]  [Supplement]  [YouTube]  [1-min Teaser]  [Poster]  [Slides]  [Code]  [BibTex] 

A new differentiable sphere tracing algorithm enables systems to efficiently render various 2D observations from deep implicit signed distance functions and effectively perform robust inverse geometric optimization over real-world images with great generalization capability.

Towards Better Generalization: Joint Depth-Pose Learning without PoseNet

, Shaohui Liu, ,

CVPR, 2020.

[arXiv]  [Supplement]  [Slides]  [Code]  [BibTex] 

Instead of relying on a PoseNet-like architecture, explicitly solving relative pose from optical flow correspondence improves the performance and generalization of self-supervised joint depth-pose learning methods on multiple challenging scenarios.

RepPoints: Point Set Representation for Object Detection

*, Shaohui Liu*, , ,

ICCV, 2019.         Intern project at .

[arXiv]  [Supplement]  [Slides]  [Blog]  [Code]  [BibTex] 

Using a set of representative points to connect stages for simultaneous semantically aligned feature extraction and flexible geometric 2D representation produces a brand new effective object recognition framework without the need of anchors and bounding boxes.

Normalized Diversification

Shaohui Liu*, *, ,

CVPR, 2019.        

[arXiv]  [Supplement]  [Slides]  [Code]  [BibTex] 

A novel loss term computed over the normalized pairwise distance matrices of the latent vectors and the corresponding outputs enforces both active extrapolation and safer interpolation of the mapping, ameliorating the notorious mode collapse problem on various vision applications.

Project August: Efficient Face Tracking at More than 1k FPS on CPU

Major project developer at in 2017.

[Project]  [YouTube] 

By caching, reusing and sharing intermediate features of a lightweight regression-based tracking model across frames we achieved extremely efficient face tracking and deployed the system onto real-world market products, as a preprocessing step for subsequent face analysis.