Christoph Feichtenhofer

Career

Christoph Feichtenhofer completed a BSc, an MSc and a PhD in computer science at TU Graz, with a doctoral thesis titled "Deep Learning for Video Recognition" (2014–2017). During those years Feichtenhofer was a visiting researcher at the York University Vision Lab with Richard P. Wildes and at the Visual Geometry Group of the University of Oxford with Andrew Zisserman, then held a postdoctoral position at TU Graz.¹²

Feichtenhofer joined Facebook AI Research (FAIR) as a research scientist in 2018. Since 2021 Feichtenhofer has worked at Meta as a research scientist and director, a role the homepage now lists under Meta Superintelligence Labs.¹²

Research

Feichtenhofer's early work centred on representations for video understanding, including two-stream network fusion for action recognition (CVPR 2016), the SlowFast architecture (ICCV 2019) and the efficient X3D family of video models (CVPR 2020). Later work extended transformers to video and images through Multiscale Vision Transformers (ICCV 2021) and the Hiera hierarchical vision transformer (ICML 2023).¹⁶

More recent papers address self-supervised and multimodal learning, such as masked autoencoders for spatiotemporal data, the MetaCLIP study of CLIP training data, and the Perception Encoder and PerceptionLM releases of 2025. Feichtenhofer is also a co-author of ConvNeXt, which revisited convolutional networks for image recognition.¹⁸

Segment Anything

Feichtenhofer is a co-author of SAM 2, which extended promptable segmentation from images to video, and of SAM 3, which adds segmentation from short noun phrases and image exemplars. The SAM 3 repository names Feichtenhofer as one of the project leads, and the homepage lists SAM 3 at ICLR 2026 and SAM Audio at ICML 2026.³⁵⁴¹

Work and contributions

• Segment Anything (SAM 2 and SAM 3) Open weights

  Co-author of the SAM 2 paper (2024) and an author of the SAM 3 paper (2025); the SAM 3 repository README names Feichtenhofer as one of the project leads.³⁵⁴

• SlowFast and PySlowFast Open-source code

  First author of "SlowFast Networks for Video Recognition" (2018); the PySlowFast README lists Feichtenhofer among the people who wrote and maintain the Apache-2.0 codebase.⁶⁷

• SAM 2.1 Hiera-Large Open weights

  Last-listed author of the SAM 2 paper (arXiv 2408.00714), which the SAM 2 repository credits; that repository released the SAM 2.1 checkpoints, including Hiera-Large, in September 2024.³⁹

• SAM 3.1 Open weights

  Last-listed author of the SAM 3 paper (arXiv 2511.16719), which the SAM 3.1 release notes point to for technical details of Object Multiplex; the SAM 3 README names Feichtenhofer a project lead.⁵¹⁰⁴

Timeline

1. 2017Completed a PhD at TU Graz with the thesis "Deep Learning for Video Recognition".¹
2. 2018Joined Facebook AI Research (FAIR) as a research scientist.¹²
3. Dec 2018Released the SlowFast Networks for Video Recognition paper as first author.⁶
4. 2021Became a research scientist and director at Meta.¹
5. Jan 2022Co-authored "A ConvNet for the 2020s", the paper introducing ConvNeXt.⁸
6. Aug 2024Co-authored the SAM 2 paper on segmenting images and videos.³
7. Nov 2025Co-authored the SAM 3 paper on segmenting anything with concepts, as a named project lead.⁵⁴

Other documented roles

Roles the sources document that are not confirmed as current.

• Research Scientist (2018–2021), Facebook AI Research (FAIR)¹

Honors and recognition

• PAMI Young Researcher Award · IEEE Computer Society Technical Committee on Pattern Analysis and Machine Intelligence, 2023¹
• ICLR Best Paper Honorable Mention (SAM 2) · International Conference on Learning Representations, 2025¹
• DOC Fellowship · Austrian Academy of Sciences²

Where they work

• Meta 106 people in the wiki

Related people

• Nikhila RaviAlso at Meta
• Matt DeitkeAlso at Meta
• Xinlei ChenAlso at Meta
• Alexander KirillovAlso at Meta
• Saining XieAlso at Meta
• Bin XiaoComputer vision · Multimodal models
• João CarreiraComputer vision · Multimodal models
• Trevor DarrellComputer vision · Multimodal models
• Ali FarhadiComputer vision · Multimodal models
• Cordelia SchmidComputer vision · Multimodal models