Chris Rockwell

I am a Ph.D. candidate in Computer Science and Engineering at the University of Michigan. I do research in computer vision and machine learning and am advised by David F. Fouhey and Justin Johnson.

Feel free to reach out! My email is cnris at umich dot edu

Github  /  Google Scholar  /  CV  /  LinkedIn  /  Twitter

Small modifications to a ViT enable computations similar to the Eight-Point algorithm, and are a good inductive bias for pose estimation.

Transformers are really good at integrating evidence across multiple views and producing a planar reconstruction.

Our forward-warping-based method enables real-time, high-quality novel view synthesis on novel objects from sparse views.

By training on both video data and 3D reconstructions in the right way, we can build models of articulations of 3D objects on ordinary video data.

Combining 3D reasoning and autoregressive modeling facilitates high-quality and consistent single-image novel view synthesis.

A simple self-training framework significantly improves 3D human pose estimation on highly-truncated Internet videos.

We increase performance of Newell et al.'s Stacked Hourglass Networks on MPII using a decoder network as attention, along with cutout and vertical flipping. In addition, we improved precision of network confidence and explored utilizing confidence for curriculum sampling of tail cases.

We improve a simple fine-tuning baseline on Meta-Dataset to within 0.1 average rank (minimum reportable difference) of the authors' best meta-learning based method using higher regularization on fine-tuning layer compared to the backbone.

I replicate and summarize Image Generation from Scene Graphs , and evaluate replicated generated images, original generated images, and ground truths using VQA.

We replicate and improve upon validation accuracy from Stacked Hourglass Networks for Human Pose Estimation using Adam and larger batch size. I led implementation of the project and it was jointly useful for research. My Pytorch implementation is published in the princeton-vl Github.

We finetune a Squeeze and Excitation ResNet classify objects appearing in road-scene images in the Driving in the Matrix dataset. With improved sampling and data augmentation, we finished top 10 in the class. I helped with implementation and improved data augmentation.

We compare several methods for link prediction on partitions of the Patent Citation Network dataset for the first time. The very sparse nature of this graph yielded low performance compared to a more typical dataset, but methods performed well in a temporal setting. I led temporal experiments and adapted representation learning method SDNE for our task.

I assisted Erik Brinkman in expanding the agent-based latency arbitrage simulation of Wah and Wellman to model a prisoner's dilemma in the high-frequency trading space.

Template