Yi Zhang

I obtained my Ph.D. at Johns Hopkins University, advised by Bloomberg Distinguished Professor Alan Yuille, where I worked on computer vision and machine learning.

I received my B.E. in EE from Tsinghua University.

Email  /  CV  /  Google Scholar  /  Github

My current research focuses on: 1) human analysis 2) generative models for (3D) computer vision, and 3) robustness and generalization.

Enabling training 3D diffusion model on large-scale noisy 3D data.

Generating images with 3D Annotations using Diffusion Models.

An analysis-by-synthesis approach for 3D human pose estimation that is highly robust to occlusions.

Animal3D consists of 3379 images collected from 40 mammal species, high-quality annotations of 26 keypoints, and importantly the pose and shape parameters of the SMAL model. We demonstrate that synthetic pre-training is a viable strategy to boost the model performance.

We explicitly reason about occlusion in multi-person 3D human pose estimation that can generalize better than using pose priors/constraints, data augmentation, or implicit reasoning.

This compositional approach allows us to reframe fine-grained recognition as zero-shot activity recognition, where a detector is composed “on the fly” from simple first-principles state machines supported by deep-learned components.

Improving model robustness to nuisance factors using "free" nuidance labels and adversarial debiasing.

A simple unified framework for failure and anomaly detection for segmantic segmentation based on a generative model and a comparison module.

We use simulated videos to augment real training data to improve model robustness to nuisance factors, e.g. novel viewpoint, change of background and human appearance.

We control hazardous factors, e.g. specularity, texturelessness and transparency, to analyze robustness of binocular stereo algorithms. Findings in virtual world is verified in real world.

Efficient online sampling algorithm for active learning from the combinatorially large synthetic data space.

An open-source tool for interacting with and obtain data from virtual worlds in Unreal Engine.

The website code is borrowed from Jon Barron's source code.