Hi! I'm Julieta :D

In 2022 I joined Reality Labs Research at Meta.

From 2018 to 2020 I was at Uber ATG Toronto, working with Raquel Urtasun. Then in 2021 I followed the team to Waabi.

In the fall of 2016 I visited Michael Black and Javier Romero in the Perceiving Systems group at MPI Tuebingen.

From 2015 to 2018 I was a PhD student in the Department of Computer Science at the University of British Columbia, supervised by Jim Little and Holger Hoos.

From 2012 to 2014 I got an MSc from the same university, working with the same nice people.

Before that I was in Mexico.

email / bluesky / github / google scholar / linkedin

Selected papers

Animation showing Sapiens model performing human-centric vision tasks including pose estimation, depth prediction, and segmentation on diverse human subjects

Foundation for human vision models

Rawal Khirodkar, Timur Bagautdinov, Julieta Martinez, Su Zhaoen, Austin James, Peter Selednik, Stuart Anderson, Shunsuke Saito
Sapiens: Foundation for Human Vision Models. In ECCV 2024 (best paper award candidate)
A very large MAE that, when fine-tuned, gets state of the art performance on a wide range of human vision tasks.

.bib
Diagram showing a sparse permutation matrix P multiplied by a colorful weight matrix W to produce a reordered matrix PW, illustrating how network weights can be permuted for better compression

Neural network compression

Julieta Martinez*, Jashan Shewakramani*, Ting Wei Liu*, Ioan Andrei Bârsan, Wenyuan Zeng, and Raquel Urtasun.
Permute, Quantize, and Fine-tune: Efficient Compression of Neural Networks. In CVPR 2021 (oral presentation)
We search for functionally-equivalent, yet easier to compress networks to achieve state-of-the-art memory-to-accuracy tradeoffs in image classification and object detection.

.bib
Neural network architecture diagram with two branches: localization using LiDAR and map intensity with cross-correlation, and perception-prediction using voxelized LiDAR with a GNN for multimodal trajectory forecasting

Joint localization, perception, and prediction

John Phillips, Julieta Martinez, Ioan Andrei Bârsan, Sergio Casas, Abbas Sadat, and Raquel Urtasun.
Deep Multi-Task Learning for Joint Localization, Perception, and Prediction. In CVPR 2021
We design an architecture that jointly performs vehicle ego-localization, object detection, and motion forecasting.

.bib
Visualization of Pit30M dataset: map of Pittsburgh showing driving routes, grid of geographic coverage tiles, and timeline chart showing data collection density across months of the year

Pit30M

Julieta Martinez, Sasha Doubov, Jack Fan, Ioan Andrei Bârsan, Shenlong Wang, Gellért Máttyus, and Raquel Urtasun.
Pit30M: A Benchmark for Global Localization in the Age of Self-Driving Cars. In ICRA 2020 (best application paper runner-up)
We propose a dataset of 30 million images and LiDAR pairs to benchmark localization at city scale.

.bib
Pipeline diagram showing neural network encoding maps into binarized representations, then compressing via Huffman and run-length encoding into binary strings, which are decoded back for localization

Low-memory localization

Xinkai Wei*, Ioan Andrei Bârsan*, Shenlong Wang*, Julieta Martinez, and Raquel Urtasun.
Learning to Localize through Compressed Binary Maps. In CVPR 2019
We train neural networks to learn very low-memory maps useful for localization in self-driving.

.bib

Faster and more accurate compressed nearest neighbour search

Julieta Martinez, Shobhit Zakhmi, Holger H. Hoos and James J. Little.
LSQ++: lower running time and higher recall in multi-codebook quantization. In ECCV 2018
We benchmark multi-codebook quantization (MCQ) approaches on an equal footing and propose two improvements that make MCQ faster and more accurate.

.bib
Two examples of 3D pose estimation: photos of roller skating and ballroom dancing with overlaid 2D skeleton detections, corresponding 2D stick figures, and reconstructed 3D poses viewed from an angle

3d pose baseline

Julieta Martinez, Rayat Hossain, Javier Romero and James J. Little.
A simple yet effective baseline for 3d human pose estimation. In ICCV 2017 (28.9% acceptance rate)
We propose a simple deep learning baseline for 3d human pose estimation that outperforms the state of the art.

.bib
Four rows of walking human stick figures showing motion prediction: gray skeletons represent ground truth past poses, red skeletons show predicted future motion extending into the sequence

Human motion prediction

Julieta Martinez, Michael J. Black, Javier Romero.
On human motion prediction using recurrent neural networks. In CVPR 2017 (29.84% acceptance rate)
We take a close look at deep recurrent approaches for human motion prediction, and propose a simple and scalable architecture that outperforms the state of the art.

.bib
Mathematical formulation of additive quantization showing the optimization objective: minimize the squared error between data matrix X and its approximation CB, where C is a codebook and B is a binary assignment matrix

Compressed nearest neighbour search

Julieta Martinez, Joris Clement, Holger H. Hoos, James J. Little.
Revisiting additive quantization. In ECCV 2016 (26.6% acceptance rate)
Additive quantization (AQ) is a promising vector compression approach for large-scale approximate nearest neighbour search. We introduce an optimization method for AQ that pushes it beyond the state of the art.

.bib
Pipeline for cross-view action recognition: motion capture database converted to trajectory features, matched against video frames of a person sitting, with pose parameters extracted for classification

3d pose from motion

Ankur Gupta*, Julieta Martinez*, James J. Little, Robert J. Woodham.
3D pose from motion for cross-view action recognition. In CVPR 2014 (29.88% acceptance rate)
An approach to improving cross-view action recognition by retrieving mocap given video sequences.

.bib

Misc

DeepViz interface showing four query images (washing machines and electrical outlets) with similarity score curves and ranked rows of retrieved similar images including appliances and face-like objects

DeepViz

My information visualization final project, taught by the wonderful Tamara Munzner. A javascript visualization tool for image retrieval (2015).


Photo of a brown donkey grazing in a green grassy field with rolling hills in the background, used as sample input for texture synthesis

Efros and Leung JS

A javascript implementation of a classic method for texture synthesis (2015).


Prisoner's Dilemma payoff matrix showing two players with probability parameters, displaying outcomes for cooperation and defection combinations (CC, CD, DC, DD)

Pavlov is no simpleton

I tried to reproduce the results of a 1993 paper on evolutionary dynamics by Sigmund and Novak. I also wrote a blog post about it (2014).



rand()

I am/have served as a reviewer for CVPR, ECCV, ICCV, IROS, ICRA, NeurIPS, AAAI, IJCAI, CVIU, and TPAMI.

I have done science outreach with GIRLsmarts4tech and UBC women in science.

On the fall of 2014 I started organizing CVRG, the Computer Vision Reading Group at UBC.