Guillem Brasó

I am a second year PhD candidate working with Laura Leal-Taixé at the Dynamic Vision and Learning group, TU Munich.

Before starting my PhD, I obtained a Master's in Mathematics from TU Munich, while working on Multi-Object Tracking at my current lab. Before that, I got a Bachelor in Mathematics from the University of Barcelona. During my Bsc, I worked with Jordi Vitrià on explainable machine learning.

In my free time, I like to climb, play the piano, and cook.

Email  /  CV  /  Google Scholar  /  Twitter  /  Github

I have a general interest in Machine Learning and Computer Vision. Some of the tasks I focus on are detection, segmentation, tracking and human pose estimation. I am also broadly interested in leveraging ideas from classical graph-based approaches and optimization in combination with deep learning to solve vision problems.

We build upon our neural solver for multi-object tracking to develop a joint model for tracking and segmentation with neural message passing. Our main contribution is an attention-based message passing framework that operates over segmentation masks and allows us to exploit synergies between data association and segmentation and SOTA results across multiple datasets.

We show that augmenting modern classification architectures with a few fully connected layers during training significantly improves their generalization performance on low-data regimes. We propose a self-distillation mechanism that allows us to keep the number of parameter constant at test-time, while yielding universal improvements across backbones in datasets.

We propose an appearance-free 3D multi-object tracker that learns to associate 3D bounding boxes over time based on their relative geometric features and spatio-temporal relationships. Our model is light-weight and generalizes remarkably well among different locations and dataset.

We propose an multi-head attention-based framework for end-to-end bottom-up multi-person pose estimation. Our main idea is to detect both keypoints and object centers, and use cross-attention among them to group keypoints into human poses.

We introduce MOTSynth, a large-scale synthetic dataset for pedestrian tracking, detection, segmentation and 2D/3D keypoint estimation. We show that trackers and detectors trained on our dataset achieve impressive generalization results when tested on real data.

Inspired by classical graph-based multi-object tracking methods, we propose a neural message passing framework for data association in multi-object tracking. We present the first graph-structured deep learning module for tracking, and achieve SOTA results on the MOTChallenge benchmarks.

This website's source code is from Jon Barron.