Terrance DeVries

I am a research scientist at Luma AI, where I work on machine learning and computer vision for 3D capture.

I received my PhD from the University of Guelph, where I was advised by Graham Taylor. I interned with Josh Susskind's research group at Apple from summer 2020 to fall 2021. In winter 2019 I worked with Michal Drodzal as an intern at Facebook AI Research, and in fall 2018 I interned with Laurens van der Maaten, also at Facebook AI Research. I did my undergad in mechanical engineering, at the University of Guelph.

Email  /  GitHub  /  Google Scholar  /  LinkedIn

Research

I'm interested in computer vision and machine learning, especially image generation and 3D scene synthesis. My goal is to create models that can generate photorealistic and explorable 3D worlds. Representative papers are highlighted.

Unconstrained Scene Generation with Locally Conditioned Radiance Fields

Generative model for learning radiance fields of realistic scenes. Allows for generating 3D scenes and exploring them with a freely moving camera.

The GIST and RIST of Iterative Self-Training for Semi-Supervised Segmentation

Alternating between training on human labeled data and psuedo-labeled data stabilizes iterative self-training, enabling significant improvement in semi-supervised segmentation quality.

Building LEGO Using Deep Generative Models of Graphs

Generative model for sequential assembly, demonstrated by building LEGO structures piece-by-piece.

Instance Selection for GANs

Removing outliers from the training set trades GAN sample diversity for higher quality samples, faster training, and reduced model capacity requirements.

ProxyNCA++: Revisiting and Revitalizing Proxy Neighborhood Component Analysis

ProxyNCA surpasses state-of-the-art metric learning methods when updated with a bag of tricks.

On the Evaluation of Conditional GANs

Conditional generative models can be evaluated by measuring the distance between real and generated joint distributions of images and conditioning.

Does Object Recognition Work for Everyone?

Publicly available object-recognition systems perform relatively poorly on household items in countries with a low household income.

Leveraging Uncertainty Estimates for Predicting Segmentation Quality

Uncertainty estimates from segmentation networks can be used to predict the quality of the segmentations.

Learning Confidence for Out-of-Distribution Detection in Neural Networks

Training a confidence estimation branch on classification networks enables identification of out-of-distribution examples.

Improved Regularization of Convolutional Neural Networks with Cutout

Randomly masking square regions of pixels from training images is an effective form of data augmentation.

Dataset Augmentation in Feature Space

Datasets can be augmented by interpolating or extrapolating between datapoints in feature space to improve classification accuracy.

Multi-task Learning of Facial Landmarks and Expression

Facial expression recognition accuracy can be improved if the model is also trained to identify facial landmarks.

Design and source code from Leonid Keselman's website, which is based on Jon Barron's website