Alex Stergiou and Ronald Poppe‘s work on AdaPool, a high-performance pooling operator with minimal additional overhead, has been published in IEEE Transactions on Image Processing (open access). By retaining more details in the downsampling, we systematically demonstrate improvements for image classification, object detection and instance segmentation.
Interestingly, our method can be inverted to upsample images. Again, we show favorable results. Code is available through Github.
A group of researchers of the UMC Utrecht (headed by Jeroen Dudink) and University Utrecht (Ronald Poppe), won the Ureka Mega Challenge 2022 with their project on automated detection of sleep in prematurely born children. In the I-See-U project, we aim to improve sleeping in this very vulnerable group, to reduce discomfort. The Ureka reward will be used to systematically validate the performance of our system.
After selection by the general audience, the final presentation was delivered by Jeroen Dudink and Chanel Sam. The jury, consisting of professionals, praised the project for its relevance, outlook and team.
Our paper “A Personal Model of Trumpery: Linguistic Deception Detection in a Real-World High-Stakes Setting” on deception detection from Donald Trump’s tweets is published in Psychological Science. Authors are Sophie van der Zee, Alice Havrileck, Aurélien Baillon and me. It’s the first (linguistic) deception model that is tailored to an individual. We achieved a 74% accuracy by considering only the use of word types (not content), obtained using LIWC. Moreover, by extensive comparisons, we demonstrate that a personal approach outperforms all known (to us) empirical and theoretical prediction models based on LIWC that have been proposed previously in literature. Our website provides a nice, interactive read: https://apersonalmodeloftrumpery.com/
Alex Stergiou has compiled a dataset of public web videos that can be obtained from GitHub. The dataset features a range of recording settings, resolutions and frame rates. As such, it can be used to benchmark frame and video interpolation techniques, as well as up- and down-sampling operators. It’s first use is in our novel pooling operator adaPool.
With Alex Stergiou and Grigorios Kalliatakis, we’ve proposed and evaluated a novel pooling operator that adds a couple of percentage points to your classification (and detection) performance with negligible computational overhead. The pooling operator, SoftPool, can easily replace MaxPool or Average Pool in any CNN. Moreover, the operation is invertible so it can also be used for image upsampling. The paper was presented at ICCV 2021 and is available on GitHub.
Already on September 27, Alex Stergiou earned his “dr.” title. Congratulations! With an impressive thesis (Efficient Modelling Across Time of Human Actions and Interactions) and a solid defense, it’s no wonder he immediately secured a research position in Prof. Dima Damen‘s lab at the University of Bristol. We’re looking forward to more great research. All the best!
On June 10, Georgios Kapidis successfully defended his thesis “A Modular Approach for the Detection and Interconnection of Objects, Hands, Locations, and Actions for Egocentric Video Understanding“. Congratulations to the young dr.! Thanks to all committee members. It was my pleasure to supervise Georgios with prof. Remco Veltkamp over the past years!
Georgios Kapidis‘ work on multi-dataset multitask learning has been accepted to IEEE Transactions on Pattern Analysis and Machine Intelligence! Congratulations! We propose a method to improve egocentric action recognition using additional datasets with potentially different label sets, and different classification and regression tasks. We can thus learn by considering only weakly related data and tasks.
The paper will be available soon. A preliminary version of the research was presented at the EPIC workshop at ICCV 2019.
Alex Stergiou‘s paper “Learn to cycle: Time-consistent feature discovery for action recognition” now appears in Pattern Recognition Letters. We present a new method that can flexibly extract spatio-temporal patterns with varying temporal scales. We show state-of-the-art performance on benchmark datasets, including Kinetics-700, Moments in Time and HACS. The code is available!
Our lie detection research with Sophie van der Zee has been nominated for the Klokhuis science award. Klokhuis was that kids program that made you become curious about science. We are honored that our research on lying in families has been nominated for their 2020 science award. Everyone can vote for any of the 10 finalists. An episode about the topic of the winner will be made and aired. Cast your vote here! (We are “Kun jij beter liegen dan je ouders?”)
Thanks to Daan van Alkemade for the 8-second teaser video!
Ronald Poppe 