
Marcus Klasson
I work as a researcher at Ericsson in the Sensing and Perception team led by José Araújo.
I obtained my PhD from KTH, where I was supervised by Hedvig Kjellström and Cheng Zhang.
After the PhD, I was a postdoc at Aalto University working with Arno Solin and Juho Kannala on uncertainty-aware methods for computer vision topics such as NeRF/Gaussian Splatting and Vision Language Models.
Research
I am interested in computer vision, deep learning, various approaches to uncertainty estimation, and their real-world applications where methods must adapt fast under limited supervision and resources.
Here is a list of my research publications:

Post-hoc Probabilistic Vision-Language Models
Anton Baumann, Rui Li, Marcus Klasson, Santeri Mentu, Shyamgopal Karthik, Zeynep Akata, Arno Solin, Martin Trapp
preprint
The Laplace approximation on top of VLMs enables analytic uncertainty propagation without additional training useful for downstream tasks and in active learning.
Anton Baumann, Rui Li, Marcus Klasson, Santeri Mentu, Shyamgopal Karthik, Zeynep Akata, Arno Solin, Martin Trapp
preprint
The Laplace approximation on top of VLMs enables analytic uncertainty propagation without additional training useful for downstream tasks and in active learning.

DeSplat: Decomposed Gaussian Splatting for Distractor-Free Rendering
Yihao Wang, Marcus Klasson, Matias Turkulainen, Shuzhe Wang, Juho Kannala, Arno Solin
CVPR, 2025
Reconstructing distractors and multi-view inconsistencies with view-specific Gaussians enables explicit separation of occluders and the underlying static 3D scene.
Yihao Wang, Marcus Klasson, Matias Turkulainen, Shuzhe Wang, Juho Kannala, Arno Solin
CVPR, 2025
Reconstructing distractors and multi-view inconsistencies with view-specific Gaussians enables explicit separation of occluders and the underlying static 3D scene.

Streamlining Prediction in Bayesian Deep Learning
Rui Li, Marcus Klasson, Arno Solin, Martin Trapp
ICLR, 2025
Local linearization of activation functions and local Gaussian approximations at linear layers bring accurate and fast predictive uncertainties scalable to ViT and GPT-2.
Rui Li, Marcus Klasson, Arno Solin, Martin Trapp
ICLR, 2025
Local linearization of activation functions and local Gaussian approximations at linear layers bring accurate and fast predictive uncertainties scalable to ViT and GPT-2.

Flatness Improves Backbone Generalisation in Few-shot Classification
Rui Li, Martin Trapp, Marcus Klasson, Arno Solin
WACV, 2025 (Oral Presentation)
We show that backbone training and selection utilizing flatness-aware training and fine-tuning can outperform previous SotA methods in multi-domain few-show classification.
Rui Li, Martin Trapp, Marcus Klasson, Arno Solin
WACV, 2025 (Oral Presentation)
We show that backbone training and selection utilizing flatness-aware training and fine-tuning can outperform previous SotA methods in multi-domain few-show classification.

Sources of Uncertainty in 3D Scene Reconstruction
Marcus Klasson, Riccardo Mereu, Juho Kannala, Arno Solin
ECCV Workshop on Uncertainty for Computer Vision, 2024
We categorized sources of uncertainties in 3D scene reconstruction using NeRF/Gaussian Splatting and proposed experimental setups for evaluating their impact.
Marcus Klasson, Riccardo Mereu, Juho Kannala, Arno Solin
ECCV Workshop on Uncertainty for Computer Vision, 2024
We categorized sources of uncertainties in 3D scene reconstruction using NeRF/Gaussian Splatting and proposed experimental setups for evaluating their impact.

Learn the Time to Learn: Replay Scheduling in Continual Learning
Marcus Klasson, Hedvig Kjellström, Cheng Zhang
TMLR, 2023
Learning schedules over which tasks to replay at different times in continual learning can outperform replaying all tasks equally or using heuristic scheduling rules.
Marcus Klasson, Hedvig Kjellström, Cheng Zhang
TMLR, 2023
Learning schedules over which tasks to replay at different times in continual learning can outperform replaying all tasks equally or using heuristic scheduling rules.

Using Variational Multi-view Learning for Classification of Grocery Items
Marcus Klasson, Cheng Zhang, Hedvig Kjellström
Patterns, 2020
Using a VAE for fusing natural images with web-scraped images and text descriptions of groceries yields more accurate classifiers compared to training with natural images only.
Marcus Klasson, Cheng Zhang, Hedvig Kjellström
Patterns, 2020
Using a VAE for fusing natural images with web-scraped images and text descriptions of groceries yields more accurate classifiers compared to training with natural images only.

A Hierarchical Grocery Store Image Dataset with Visual and Semantic Labels
Marcus Klasson, Cheng Zhang, Hedvig Kjellström
WACV, 2019
Dataset for grocery item classification with natural images from grocery stores organized with hierarchical labels, where each class has a corresponding web-scraped text and iconic image.
Marcus Klasson, Cheng Zhang, Hedvig Kjellström
WACV, 2019
Dataset for grocery item classification with natural images from grocery stores organized with hierarchical labels, where each class has a corresponding web-scraped text and iconic image.