Dian Chen · 陈典

I'm a Research Engineer at Toyota Research Institute in the Large Behavior Model team led by Dr. Russ Tedrake, where I work on 3D/video foundation models. Before that I was a researcher at Prof. Trevor Darrell's lab at UC Berkeley EECS, where I worked on domain adaptation.

Prior to my work I obtained my master's degree in Robotics at the University of Pennsylvania, GRASP Lab, advised by Prof. Kostas Daniilidis. Before that I did my undergrad at Zhejiang University, by the beautiful West Lake in Hangzhou.

News

March 2025: MVGD accepted to CVPR 2025!
January 2025: Omnishape accepted to ICRA 2025!
September 2024: EPIO accepted to NeurIPS 2024!
Feburary 2024: pix2gestalt accepted to CVPR 2024, as an Oral presentation!
January 2024: FSD accepted to ICRA 2024!
July 2023: ZeroDepth accepted to ICCV 2023!
June 2023: I will co-organize and give an oral talk at the VCAD workshop at CVPR 2023.
March 2023: VEDet and PF-Track accepted to CVPR 2023!
January 2023: DD3Dv2 accepted to ICRA 2023!
March 2022: AdaContrast and DepthFormer accepted to CVPR 2022!

Research

Zero-Shot Novel View and Depth Synthesis with Multi-View Geometric Diffusion CVPR 2025
Conference on Computer Vision and Pattern Recognition (CVPR), 2025
- 3D geometry
- diffusion
- generalization
- foundation model
Multi-view Geometric Diffusion (MVGD) is a framework for zero-shot images and scale-consistent depth maps generation from novel viewpoints given an arbitrary number of posed input views.
OmniShape: Zero-Shot Multi-Hypothesis Object Pose and Shape Estimation in the Real World ICRA 2025
International Conference on Robotics and Automation (ICRA), 2025
- 3D geometry
- diffusion
- pose & shape estimation
OmniShape estimates object pose and full shape via two conditional diffusion models, enabling multi-hypothesis sampling from the joint distribution without assuming known 3D models.
- [paper]
- [project page]
SE(3) Equivariant Ray Embeddings for Implicit Multi-View Depth Estimation Neurips 2024
Conference on Neural Information Processing Systems (NeurIPS), 2024
- 3D geometry
- depth estimation
We propose to embed SE(3) equivariance into the Perceiver IO architecture to make geometric learning more data-efficient with geometry integrity.
- [paper]
pix2gestalt: Amodal Segmentation by Synthesizing Wholes CVPR 2024
Conference on Computer Vision and Pattern Recognition (CVPR), 2024
- amodal segmentation
- diffusion
- generalization
pix2gestalt synthesizes whole objects from only partial visible ones, enabling amodal segmentation, recognition, and 3D reconstruction of occluded objects.
FSD: Fast Self-Supervised Single RGB-D to Categorical 3D Objects ICRA 2024
International Conference on Robotics and Automation (ICRA), 2024
- self-supervision
- 3D geometry
- 6D pose estimation
A framework for fast self-supervised categorical 6D pose and size estimation and shape reconstruction.
Towards Zero-Shot Scale-Aware Monocular Depth Estimation ICCV 2023
International Conference on Computer Vision (ICCV), 2023
- 3D geometry
- depth estimation
- generalization
- foundation model
A step toward a generalizable 3D foundation depth model, which predicts metric depth for arbitrary test images from different domains and camera parameters.
Viewpoint Equivariance for Multi-View 3D Object Detection CVPR 2023
Conference on Computer Vision and Pattern Recognition (CVPR), 2023
- 3D geometry
- object detection
Improve multi-view 3D object detection by injecting 3D geometry as input-level inductive bias to a transformer, and regularizing the multi-view consistency with viewpoint equivariance loss.

^{Featured in oral talk in Vision-Centric Autonomous Driving (VCAD) Workshop at CVPR 2023!}
Standing Between Past and Future: Spatio-Temporal Modeling for Multi-Camera 3D Multi-Object Tracking CVPR 2023
Conference on Computer Vision and Pattern Recognition (CVPR), 2023
- temporal learning
- multi-object tracking
An end-to-end tracking framework that reasons from the "past" and makes predictions into the "future", featuring drastically lower ID-switches.
Depth Is All You Need for Monocular 3d Detection ICRA 2023
International Conference on Robotics and Automation (ICRA), 2023
- self-supervision
- 3D geometry
- object detection
- representation learning
Improve 3D monocular perception by enhancing the model's 3D representation by self-supervised depth pre-training.
- [paper]
- [code]
Contrastive Test-Time Adaptation CVPR 2022
Conference on Computer Vision and Pattern Recognition (CVPR), 2022
- self-supervision
- domain adaptation
- representation learning
Exploit self-supervised contrastive learning to adapt the model's representation at test-time upon distribution shifts.

^{Featured in keynote talk in Robustness in Sequential Data Workshop at CVPR 2022!}
Multi-Frame Self-Supervised Depth with Transformers CVPR 2022
Conference on Computer Vision and Pattern Recognition (CVPR), 2022
- self-supervision
- 3D geometry
- depth estimation
Leverage geometric relationships between frames and attention mechanism, to learn powerful feature matching for points sampled on epipolar lines using transformers.
- [paper]
- [code]
Region-Level Active Learning for Cluttered Scenes Arxiv 2021
- active learning
- object detection
Combine the best of both image-level and object-level approach into a generalized, region-level active learning framework for satellite image object detection.
- [paper]
Minimax Active Learning Arxiv 2021
- active learning
- representation learning
Active learning via a minimax game on classification entropy.
- [paper]

Service

Conference & Journal reviewer: CVPR (2022, 2023, 2024), ICCV/ECCV (2023, 2024), IJCV (2024)
Mentor in Women in Computer Vision Workshop at CVPR 2023

Maze of the day