Philipp Krähenbühl

Also known as Philipp Kraehenbuehl, Philipp Krahenbuhl

Philipp Krähenbühl is a computer vision researcher at the University of Texas at Austin whose work includes fully connected CRF inference, Context Encoders and the CenterNet object detector. He took first place at the Swiss Olympiad in Informatics in 2004 and 2005 and won a silver medal at the International Olympiad in Informatics in 2005. With Vladlen Koltun, he received a NIPS Outstanding Student Paper Award in 2011, and he earned his Stanford PhD in 2014, followed by two years as a postdoc at UC Berkeley. In 2016, he coauthored Context Encoders and joined UT Austin as an assistant professor; he received an NSF CAREER Award in 2019 and coauthored 'Objects as Points', which introduced CenterNet. He is an associate professor of computer science at UT Austin and a researcher at Apple, working on multimodal foundation models, interactive agents and reinforcement learning post-training of large models.

Now at: University of Texas at Austin and Apple
Badges: Bridge

Timeline

2004–2005

Swiss and international Olympiads

He takes first place at the Swiss Olympiad in Informatics in 2004 and 2005. In 2005, he also wins a silver medal at the International Olympiad in Informatics and another at the Swiss Mathematics Olympiad.
2011

A NIPS student paper award

He and Vladlen Koltun receive a NIPS Outstanding Student Paper Award in 2011 for 'Efficient Inference in Fully Connected CRFs with Gaussian Edge Potentials'. The paper's algorithm approximately minimizes fully connected models on tens of thousands of variables in a fraction of a second.
2014–2016

Stanford PhD, Berkeley postdoc

He receives his PhD from Stanford's computer science department in 2014 and then spends two years as a postdoc at UC Berkeley.
2016

Context Encoders for inpainting

He coauthors Context Encoders in 2016 with Deepak Pathak, Jeff Donahue, Trevor Darrell and Alexei A. Efros. The convolutional neural network learns to generate the contents of an arbitrary image region from its surroundings, and its learned features serve CNN pre-training and semantic inpainting.
2016

Assistant professor at UT Austin

He joins the University of Texas at Austin's computer science department as an assistant professor in 2016. He studies computer vision, with a focus on reducing the number of labels needed to teach computers what to see.
2019

An NSF CAREER Award

He receives a National Science Foundation CAREER Award in 2019, the NSF's award for early-career faculty. He plans to develop efficient, robust and compact video representations, which can reduce the economic and environmental costs of digital infrastructure.
2019

Objects as Points and CenterNet

In 2019, he coauthors 'Objects as Points' with Xingyi Zhou and Dequan Wang. The paper models an object as a single point, the center of its bounding box, and its CenterNet detector reaches 28.1% AP at 142 FPS on the MS COCO dataset.
As of October 2026

UT Austin professor, Apple researcher

He is an associate professor in the Department of Computer Science at the University of Texas at Austin and a researcher at Apple, with interests in multimodal foundation models, interactive agents and reinforcement learning post-training of large models. His ICLR 2025 paper 'Cut Your Losses in Large-Vocabulary Language Models', written with Erik Wijmans, Brody Huval, Alexander Hertzberg and Vladlen Koltun, introduces Cut Cross-Entropy, which cuts the memory footprint of the loss computation for Gemma 2 (2B) from 24 GB to 1 MB.