Yunlu Chen

Ph.D. candidate, University of Amsterdam

3D deep learning
Google Scholar

Contact: ychen9201 AT gmail.com

Bio

I am a final-year PhD student at University of Amsterdam (UvA), working on 3D deep learning.

I am grateful to be supervised by Efstratios Gavves, Thomas Mensink and Arnold Smeulders at UvA. I also received guidance from some great external senior researchers: Basura Fernando (A*STAR, Singapore), Hakan Bilen (University of Edinburgh), and Matthias Nießner (Technical University of Munich) that collaborated in my research projects.

Prior to my Ph.D., I obtained my B.Sc. in Physics from Fudan University, China, and my M.Sc. (cum laude) in Artificial Intelligence from Utrecht University, the Netherlands.

Research

My research lies at the intersection of deep learning and 3D computer vision. I build deep learning models for different 3D modalities. Some prior projects include monocular depth estimation, RGB-D semantic segmentation, and augmentation for point cloud recognition.

My recent projects focus on the fascinating field of neural implicit 3D representations, with the sense of continuity in the 3D Euclidean space that we live in. Some interesting highlights of my research in this direction are the observation of the emerging layer hierarchy and correspondence in the latent-coded implicit functions, as well as the improved generalization to unseen similarity transformations by leveraging the graph embedding and the inductive prior of equivariance.

Papers

3D Equivariant Graph Implicit Functions
Yunlu Chen, Basura Fernando, Hakan Bilen, Matthias Nießner, Efstratios Gavves

ECCV 2022

[paper] [Project] [video]

We propose a novel family of graph-based 3D implicit representations. The non-Euclidean graph embeddings in multiple-scales enable modeling of high-fidelity 3D geometric details and facilitate inherent robustness against geometric transformations. We further incorporate equivariant graph layers for guaranteed generalization to unseen similarity transformations, including rotation, translation, and scaling.

Neural Feature Matching in Implicit 3D Representations
Yunlu Chen, Basura Fernando, Hakan Bilen, Thomas Mensink, Efstratios Gavves

ICML 2021

[paper] [supp]

We observe the emerging structure in implicit representation model layers: early layers reconstruct the coarse shape outline, and deeper layers refine finer details. We additionally show that hidden features in implicit representations learns high-level semantic concepts without supervision, which further allows applications for point matching and deformation between shapes with different semantic components and topology.

Unsharp Mask Guided Filtering
Zenglin Shi, Yunlu Chen, Efstratios Gavves, Pascal Mettes, Cees G.M. Snoek

TIP 2021

[paper] [code]

We propose a guided image filtering network that predicts high-frequency residuals from the low-pass filtered input signals, resulting in state-of-the-art results across filtering tasks like upsampling, denoising, and cross-modality filtering.

PointMixup: Augmentation for Point Cloud
Yunlu Chen, Vincent Tao Hu, Efstratios Gavves, Thomas Mensink, Pascal Mettes, Pengwan Yang, Cees G.M. Snoek

ECCV 2020 (spotlight)

[paper] [supp] [code]

We developed PointMixup, a novel augmentation and regularization for point clouds. We analyse the mathematical properties of the shortest-path interpolation wrt Earth Mover's Distance, which explains the effectiveness of the proposed method.

3D Neighborhood Convolution: Learning Depth-Aware Features for RGB-D and RGB Semantic Segmentation
Yunlu Chen, Thomas Mensink, Efstratios Gavves

3DV 2019

[paper]

We incorporate depth channel information into 2D network without using a separate stream. Depth information is used to adjust the receptive field of the 2D convolution kernel to be similar to a local 3D neighbourhood.