Since 2021, I’ve been a research scientist at Waabi, where I currently work on simulation. My research focuses on bringing generative AI and neural rendering into the physical world through mixed reality. The aim is to improve autonomy and devise safe and effective ways of testing self-driving vehicles in challenging and safety-critical scenarios. We are hiring!

Before this, I spent three wonderful years as a researcher at Uber Advanced Technologies Group (ATG) Toronto, working on applying my research to the challenges associated with autonomous driving in the real world—from state estimation to perception and prediction.

In October 2024 I completed my PhD in the Machine Learning Group at the University of Toronto under the supervision of Professor Raquel Urtasun. My PhD research focused on computer vision and deep learning for robotics and long-term autonomy.

Education

Publications

Siva Manivasagam*Ioan Andrei Bârsan*Jingkang WangZe Yang and Raquel Urtasun
Note: *denotes equal contribution.
Web PDF BibTeX Poster

TL;DR: We propose a thorough methodology for evaluating how realistic a self-driving vehicle simulator is, and use it to quantify the importance of reflectance modeling and other factors when simulating LiDAR.

Jingkang WangSiva ManivasagamYun ChenZe YangIoan Andrei BârsanJoyce YangWei-Chiu Ma and Raquel Urtasun
Note: *denotes equal contribution.
Web PDF BibTeX

TL;DR: Automatically build "game-ready" rigged 3D meshes from observed data by optimizing vertices and material properties from a template mesh.

John Phillips*Julieta Martinez*Ioan Andrei Bârsan*Sergio CasasAbbas Sadat and Raquel Urtasun
Note: *denotes equal contribution.
Web PDF (arXiv) BibTeX Poster Video (Download) Video (YouTube)

TL;DR: We show that object detection and prediction systems for self-driving cars can tolerate relatively large sensor-to-map misalignments (up to 0.5m) without errors increasing too much. However, motion planning is much more sensitive. We propose a lightweight 2ms-overhead multi-task approach to correct the pose and increase resilience to localization errors.


Julieta Martinez*Jashan Shewakramani*, Ting Wei Liu*Ioan Andrei BârsanWenyuan Zeng and Raquel Urtasun
Note: *denotes equal contribution.
PDF (arXiv) BibTeX Code

TL; DR: A simple yet effective low-overhead method for compressing neural network weights using a form of product quantization.

By permuting (Step 1) the rows of the weight matrices in an optimal way, we can maximize the effectiveness of quantization (Step 2). The key ingredient is fine-tuning (Step 3) the dictionary post-quantization, while keeping the weight codes fixed. This can be done using vanilla autodiff, e.g., in PyTorch.

We show this approach maintains the vast majority of the original networks' performance on classification and object detection, while reducing the memory footprint of their parameters by nearly 20x. The code is open source.

Joyce Yang*Can Cui*Ioan Andrei Bârsan*Raquel Urtasun and Shenlong Wang
Note: *denotes equal contribution.
Web PDF (arXiv) BibTeX Talk Video (YouTube) Results Video

TL;DR: We analyze Simultaneous Localization and Mapping (SLAM) in a setting where multiple cameras are attached to a robot but fire at different times, e.g., by following a spinning LiDAR. We extend a classic SLAM formulation with a continuous time motion model that integrates these asynchronous observations robustly and efficiently. Our system robustly initializes and tracks its pose in crowded environments and closes loops using all camera information.

We evaluate our method on a new large-scale multi-camera SLAM benchmark and demonstrate the benefits of asynchronous sensor modeling at scale.

Dataset overview map.
Julieta MartinezSasha DoubovJack FanIoan Andrei BârsanShenlong WangGellért Máttyus and Raquel Urtasun
International Conference on Intelligent Robots and Systems (IROS) 2020

Best Application Paper Finalist!

Web PDF (arXiv) BibTeX Play with it! Overview Video (90s) IROS Talk (15min) Code

TL;DR: A new self-driving dataset containing >30M HD images and LiDAR sweeps covering Pittsburgh over one year, all with centimeter-level pose accuracy. We investigate the potential of retrieval-based localization in this setting, and show that simple architecture (e.g., ResNet + global pool) perform surprisingly well, outperforming more complex architectures like NetVLAD.

The figure shows the geographic (top) and temporal (bottom, x = date, y = time of day) extent of the data.

Illustration.
Wei-Chiu Ma*Ignacio Tartavull*Ioan Andrei Bârsan*Shenlong Wang*Min BaiGellért MáttyusNamdar HomayounfarShrinidhi Kowshika LakshmikanthAndrei Pokrovsky and Raquel Urtasun
International Conference on Intelligent Robots and Systems (IROS) 2019
Note: *denotes equal contribution.
PDF (arXiv) BibTeX Talk Slides (PDF) Talk Slides (Apple Keynote) Video

TL;DR: We use very sparse maps consisting in lane graphs (i.e., polylines) and stored traffic sign positions to localize autonomous vehicles. These maps take up ~0.5MiB/km2, compared to, e.g., LiDAR ground intensity images which can take >100MiB/km2. We use these maps in the context of a histogram filter localizer, and show median lateral accuracy of 0.05m and median longitudinal accuracy of 1.12m on a highway dataset.

Compressed maps demo image
Xinkai Wei*Ioan Andrei Bârsan*Shenlong Wang*Julieta Martinez and Raquel Urtasun
International Conference on Computer Vision and Pattern Recognition (CVPR) 2019
Note: *denotes equal contribution.
PDF PDF (arXiv) BibTeX Poster Video

TL;DR: High-resolution maps can take up a lot of storage. We use neural networks to perform task-specific compression to address this issue by learning a special-purpose compression scheme specifically for localization. We achieve two orders of magnitude of improvement over traditional methods like WebP, as well as less than half the bitrate of a general-purpose learning-based compression scheme. For reference, PNG takes up 700× more storage on our dataset.

Localizer preview image
Ioan Andrei Bârsan*Shenlong Wang*Andrei Pokrovsky and Raquel Urtasun Proceedings of the Second Conference on Robot Learning (CoRL) 2018
Note: *denotes equal contribution.
PDF BibTeX Poster Talk Slides (PDF) Video

TL;DR: Matching-based localization methods using LiDAR can provide centimeter-level accuracy, but require careful beam intensity calibration in order to perform well. In this paper, we cast the matching problem as a learning task and show that it is possible to learn to match online LiDAR observations to a known map without calibrated intensities.

Map preview image
Ioan Andrei BârsanPeidong LiuMarc Pollefeys and Andreas Geiger
IEEE International Conference on Robotics and Automation (ICRA) 2018
Web PDF BibTeX Poster Code

TL;DR: A system for outdoor online mapping using a stereo camera capable of also reconstructing the dynamic objects it encounters, in addition to the static map. Supports map pruning to eliminate stereo artifacts and reduce memory consumption to less than half.

Work Experience

Industry

  • Full-time senior scientist at a Waabi (Mar 2021–Present).
  • Full-time research scientist at Uber ATG Toronto (Jan 2018–Feb 2021).
    • Helping develop scalable and robust centimeter-accurate localization methods for self-driving cars.
    • LiDAR-based map localization, visual localization, learning-based compression, large-scale machine learning (Apache Spark).
    • Multi-Task Learning for autonomous driving with a focus on real-time operation.
    • Data engineering; petabyte-scale data ingestion, curation, and benchmark selection.
  • Previously, I did a series of software engineering internships in the US during my undergrad:
    • Internship: Twitter (Summer 2015, San Francisco, CA), Performance Ads
      • Developed Apache Storm and Hadoop data pipelines using Scala.
    • Internship: Google (Summer 2014, New York, NY), Data Protection
      • Co-developed a system for performing security-oriented static analysis of shell scripts used to run large numbers of cluster jobs.
    • Internship: Microsoft (Summer 2013, Redmond, WA), Server and Tools Business
      • Security and reliability analysis of a web service part of the Azure portal.

Academic

  • Peer review: IJCV 2021, CVPR (2021–present), ECCV/ICCV (2020–present), NeurIPS (2022), ICLR 2022, ICRA (2019, 2021–present), IROS (2019–present), CoRL 2020, AAAI 2021, RA-L (2020–present)
    • Acknowledged as one of the top reviewers for ECCV 2020 (top 7.5%) and ECCV 2022.
    • Acknowledged as an outstanding reviewer for CVPR 2021.
    • Acknowledged as one of the top reviewers for NeurIPS 2022.
  • Teaching Assistant: Image Analysis and Understanding (CSC420), University of Toronto, Fall 2017.

Talks

Miscellaneous

Bio

Before starting my PhD, I completed my Master’s in Computer Science at ETH Zurich. For my Master’s Thesis, I developed DynSLAM, a dense mapping system capable of simultaneously reconstructing dynamic and potentially dynamic objects encountered in an environment, in addition to the background map, using just stereo input. More details can be found on the DynSLAM project page.

Previously, while doing my undergraduate studies at Transilvania University, in Brașov, Romania, I interned at Microsoft (2013, Redmond, WA), Google (2014, New York, NY) and Twitter (2015, San Francisco, CA), working on projects related to privacy, data protection, and data pipeline engineering.

I am originally from Brașov, Romania, a lovely little town which I encourage everybody to visit, together with the rest of Southeast Europe.

Contact

Email me at iab (at) cs (dawt) toronto (dawt) edu.

Find me on Twitter, GitHub, Google Scholar, LinkedIn, or StackOverflow.