PeopleofSuperintelligence

Researcher

Joao Carreira

Also known as João Carreira, Joao Luis da Silva Carreira

Joao Carreira is a computer vision researcher at Google DeepMind whose work centers on learning from video. With Cristian Sminchisescu, he developed the CPMC object segmentation method and was part of the Bonn team that took first place in the 2010 PASCAL segmentation challenge; after a 2012 PhD at the University of Bonn, he worked as a postdoc with Jitendra Malik at Berkeley. At DeepMind, he co-created the Kinetics action dataset and, with Andrew Zisserman, the I3D video model in 2017, then coauthored the Perceiver models in 2021 and the Perception Test benchmark in 2023. A paper on learning image encoders from one long video earned him and his coauthors an honorable mention among the ICLR 2024 awards, and in December 2024 he and his coauthors showed that masked video models keep improving on 4D tasks up to 22 billion parameters. Today he is a research scientist in London, and in September 2026 he co-organized the Perception Test workshop at ECCV 2026 in Malmö.

Badges

Timeline

  1. 2010

    CPMC and a PASCAL win

    With Cristian Sminchisescu, he presents Constrained Parametric Min-Cuts (CPMC) at CVPR 2010, a method that extracts object hypotheses as figure-ground segmentations and learns to rank them. In 2010, he, Sminchisescu and Fuxin Li of the Computer Vision and Machine Learning group at the University of Bonn's Institute for Numerical Simulation take first place in the PASCAL segmentation challenge, together with researchers from Barcelona. Their system scores 40 percent, ahead of Berkeley at 36 percent.

  2. 2012–2015

    Postdoc with Jitendra Malik

    He completes his PhD on object segmentation at the University of Bonn in 2012, advised by Cristian Sminchisescu. As a postdoc with Jitendra Malik at Berkeley, he works on 3D and human pose estimation, and in 2015 he coauthors Human Pose Estimation with Iterative Error Feedback with Pulkit Agrawal, Katerina Fragkiadaki and Malik. The model refines an initial pose estimate by feeding back its own error predictions.

  3. 2017–2018

    Kinetics and the I3D model

    At DeepMind, he is one of the core creators of the Kinetics human action video dataset, described in May 2017 with 400 action classes and at least 400 YouTube clips per class. With Andrew Zisserman, he introduces the Two-Stream Inflated 3D ConvNet (I3D) in 2017, which expands the filters and pooling kernels of image classification ConvNets into 3D to learn from video. In 2018, he and four coauthors extend the dataset to Kinetics-600, with 600 classes and some search queries in Portuguese.

  4. 2021

    The Perceiver models

    In March 2021, he coauthors the Perceiver (ICML 2021) with Andrew Jaegle, Oriol Vinyals, Andrew Zisserman and others, a Transformer-based model that iteratively distills very large inputs into a small latent bottleneck and works across images, point clouds, audio and video. In July 2021, he coauthors Perceiver IO, which adds a flexible querying mechanism so one architecture can handle tasks from language understanding to optical flow and StarCraft II.

  5. 2023

    The Perception Test benchmark

    As one of its core creators, he presents the Perception Test in May 2023 with Viorica Pătrăucean, Andrew Zisserman and others, a multimodal video benchmark of 11.6k real-world videos filmed by around 100 participants that probes memory, abstraction, physics and semantics. Published in the NeurIPS 2023 Datasets and Benchmarks track, it shows humans at 91.4% and state-of-the-art video question-answering models at 46.2%.

  6. 2024

    An ICLR 2024 honorable mention

    Is ImageNet worth 1 video?, his paper with Shashank Venkataramanan, Mamshad Nayeem Rizve, Yuki Asano and Yannis Avrithis, receives an honorable mention among the ICLR 2024 awards. The work introduces the Walking Tours dataset of hours-long first-person videos and the DoRA method, with which a single video becomes a strong competitor to ImageNet for several image and video tasks.

  7. December 2024

    Scaling 4D representations

    In December 2024, he is first author of Scaling 4D Representations, written with Andrew Zisserman and more than 30 other coauthors. The paper shows that masked auto-encoding with transformer video models keeps improving on camera pose estimation, point and object tracking and depth estimation as models grow from 20 million to 22 billion parameters, the largest self-supervised video model reported at that point.

  8. As of October 2026

    Research scientist at Google DeepMind

    He is a research scientist at Google DeepMind in London, where his team works on 4D reconstruction, multitask vision models, city-scale perception and modelling the future for acting and robotics. In September 2026, at ECCV 2026 in Malmö, Sweden, he co-organizes the workshop Perception Test: From Tabletop to City Scale and coauthors the papers Forecasting Motion in the Wild and Kilometer-Vision.

Career

  1. Since 2017Google DeepMindResearch scientist
  2. 2012–2015University of California, BerkeleyPostdoc

Education

  1. 2012University of BonnPhD

From the biography.

  1. The median h-index is 48 at universities and 14 at frontier AI labsIncludes Joao Carreira. The median h-index is 48 at universities and 14 at frontier AI labs. We compare publishing careers, recent citations, dual appointments and job moves across 2,420 university researchers and 308 frontier-lab researchers.
  2. Google had the most people shaping AI every year from 2014 to 2025Includes Google. Our estimates of which organizations employed the most people shaping AI each year from 1955 to 2025, tracing the shift from university labs to companies.
  3. Google has the highest combined h-index among AI organizationsIncludes Google. We estimate that Google holds the highest combined h-index at 8,016, while universities account for 73% of the deduplicated total across AI organizations.
  4. 20% of people at companies who shape AI also hold a university role, led by Google and Stanford UniversityIncludes Google. We estimate that 539 of the 6,766 people with a known current organization (8%) hold both a university role and a company role, including 20% of people at companies and 13% of people at universities.

Colleagues

People named in this biography, and people with dated positions at the same organization in the same years.

Sources and credits

Photo: We have no published photo of Joao Carreira yet. Send us a free-licensed photo.

  1. Constrained Parametric Min-Cuts for Automatic Object Segmentation | ML Anthology mlanthology.org. Checked 20 October 2026.
  2. Preisgekrönt: Bonner Computer können am besten sehen — Universität Bonn uni-bonn.de. Checked 20 October 2026.
  3. Joao Luis da Silva Carreira simons.berkeley.edu. Checked 20 October 2026.
  4. [1507.06550] Human Pose Estimation with Iterative Error Feedback arxiv.org. Checked 20 October 2026.
  5. [1705.06950] The Kinetics Human Action Video Dataset arxiv.org. Checked 20 October 2026.
  6. [1705.07750] Quo Vadis, Action Recognition? A New Model and the Kinetics Dataset arxiv.org. Checked 20 October 2026.
  7. [1808.01340] A Short Note about Kinetics-600 arxiv.org. Checked 20 October 2026.
  8. [2103.03206] Perceiver: General Perception with Iterative Attention arxiv.org. Checked 20 October 2026.
  9. [2107.14795] Perceiver IO: A General Architecture for Structured Inputs & Outputs arxiv.org. Checked 20 October 2026.
  10. [2305.13786] Perception Test: A Diagnostic Benchmark for Multimodal Video Models arxiv.org. Checked 20 October 2026.
  11. ICLR 2024 Awards iclr.cc. Checked 20 October 2026.
  12. [2412.15212] Scaling 4D Representations arxiv.org. Checked 20 October 2026.
  13. Keynote Speeches conference.cs.cityu.edu.hk. Checked 20 October 2026.
  14. Google at ECCV 2026 research.google. Checked 20 October 2026.