Embedded Files
SCIA 2027
SCIA 2027
SCIA 2027 (Scandinavian Conference on Image Analysis) will be held at NTNU Gjøvik June 9–11, 2027, and the list of keynotes has now been announced! We are thrilled to welcome:
Alexei Efros (UC Berkeley)
Aasa Feragen (DTU)
Sabine Süsstrunk (EPFL)
The conference is one of the most important gatherings for researchers and professionals in image analysis in the Nordic region, and features expert keynotes, high-quality scientific presentations, and networking opportunities with leading figures from both academia and industry.
Important dates:
Paper submission opens: 01.12.2026
Paper submission deadline: 26.01.2027
Abstract submission opens: 12.02.2027
Abstract submission deadline: 28.02.2027
Notification of acceptance: 16.03.2027
Camera-ready deadline: 11.04.2027
Early-bird registration deadline: TBA
Tutorials and workshops: 08.06.2027
Conference dates: 09.–11.06.2027
Every two years, the Best Nordic Thesis Prize is also awarded to recognize outstanding research by emerging scientists.
Check out the website for information and registration: https://scia2027.org
NOBIM Biannual Conference: Moving to Early 2028
NOBIM Biannual Conference: Moving to Early 2028
The next NOBIM biannual conference is now planned for early 2028. The conference has been shifted by one year to avoid overlap with SCIA 2027. Unfortunately, 2026 was not a viable option given the short amount of time available to organize the event. We look forward to welcoming old and new faces in early 2028, and more details will be shared as they become available.
Congratulations: Former NOBIM Leader Named IAPR Fellow
Congratulations: Former NOBIM Leader Named IAPR Fellow
We are delighted to announce that Robert Jenssen, professor at UiT The Arctic University of Norway and a former head of NOBIM, has been named an IAPR Fellow, recognized at ICPR 2026 for outstanding contributions to pattern recognition and machine learning.
In his own words:
"I have always immensely valued the IAPR for its 50-year impact on machine learning and AI via its roots in pattern recognition research. I am very happy to contribute in a new role as a Fellow. I would especially like to thank Ingela Nyström and Jianying Hu for endorsement and leadership, as well as the Norwegian IAPR Society for the nomination.
Congratulations on this well-deserved recognition!
Featured Article: Suppressing Non-Semantic Noise in Masked Image Modeling Representations
Featured Article: Suppressing Non-Semantic Noise in Masked Image Modeling Representations
By Martine Hjelkrem-Tan and Changkyu Choi
Modern computer vision increasingly relies on foundation models trained on vast amounts of unlabeled images. One of the most widely used training recipes is masked image modeling (MIM), where parts of an image are masked out and the model learns by predicting what lies behind the mask. The approach scales well and underpins many of today's powerful vision backbones.
This work shows that the recipe carries a hidden cost. To fill in a missing patch, a model benefits from knowing where that patch sits in the image, not only what it depicts. Our analysis finds that MIM-trained models devote a substantial share of their internal capacity to this positional bookkeeping, and that the effect appears consistently across architectures and training variants while being nearly absent in models trained with non-MIM recipes.
To measure the effect, we introduce a semantic invariance score. We feed the model both real images and synthetic images that carry no meaningful content, and we look for internal directions that respond the same way to both. Our claim is that a direction that cannot tell an image from noise is not carrying semantic information.
Building on this diagnostic, we propose Semantically Orthogonal Artifact Projection (SOAP), which removes the identified directions from the representation. SOAP requires no training, is computed directly from data, and attaches to any pretrained backbone as a single linear layer. Across MIM-based models it improves zero-shot object segmentation and classification without modifying the underlying weights.
Beyond the immediate gains, the work offers a diagnostic viewpoint for pretraining design, indicating where a model's capacity is being spent and how much of it actually encodes what we care about. The paper appears at CVPR 2026 in Denver, Colorado, and the code is publicly available [1].
[1] M. Hjelkrem-Tan et al., Suppressing Non-Semantic Noise in Masked Image Modeling Representations, CVPR 2026. github.com/dsb-ifi/soap
Are you our next featured article?
Are you our next featured article?
Get in touch with us at styret@nobim.no should you want your research to be featured.
Page updated
Google Sites
Report abuse