The 42nd Swedish Symposium on Image Analysis
The 8th Swedish Symposium on Deep Learning
KTH Royal Institute of Technology
Stockholm, Sweden, March 12-14, 2025
The SSBA symposium is the premier Swedish event where researchers, industrial professionals and students gather to learn about the recent developments in the areas of image processing, computer vision, pattern recognition and related fields. SSBA 2025 will feature keynote speakers, and oral presentations and posters of submitted papers.
SSDL is an important Swedish forum for leading research groups in industry and academia to meet and discuss the latest trends and developments in deep learning and related areas. SSDL 2025 will feature invited talks by leading researchers in deep learning as well as oral and poster presentations of submitted papers and abstracts.
Conference sponsors:
Award sponsor:
In this presentation, I will discuss our progress toward building models that learn autonomously from raw data such as images and videos. These models aim to restore degraded observations, construct meaningful representations, and predict future states without requiring manual guidance. Our ultimate vision is to develop systems capable of autonomously complete fundamental tasks like navigation and object interaction. This pursuit is driven by two key motivations: (1) The sheer volume of data necessitates minimal human intervention to make autonomous learning scalable and practical; and (2) By enabling models to learn directly from raw observations, we open the door to discovering novel and potentially more effective solutions, free from the biases introduced by human-designed objectives. I will present examples of our work in several key areas, including Computational Photography, Representation Learning, and Controllable World Models, showcasing the potential and challenges of building truly autonomous learning systems. Here is his homepage.
The advancement of Generative Adversarial Networks (GANs) and Diffusion Models has revolutionized our ability to generate highly realistic virtual images and videos.
These AI-generated creations, commonly referred to as Deepfakes, represent a significant leap in artificial intelligence capabilities. Deepfake technology has made
remarkable contributions, particularly in fields such as entertainment, where it is employed to create highly immersive visual effects, virtual actors, and innovative
storytelling techniques. However, the same technology has also given rise to serious ethical and societal challenges. Deepfakes are increasingly exploited for malicious
purposes, including the spread of misinformation, defamation, and personal attacks. These incidents have been widely reported, raising concerns about the misuse of AI
and its impact on trust in media and communication. Our research group has been at the forefront of exploring the dual nature of Deepfake technology. We have conducted
extensive studies on detecting and mitigating Deepfake abuse through cutting-edge Computer Vision algorithms and developing tools to ensure the responsible use of this
technology. Moreover, we are investigating the potential of Deepfakes to drive advancements in Computer Vision, such as improved data augmentation techniques, enhanced
virtual reality experiences, and more sophisticated human-computer interaction.
In this talk, I will provide an overview of our representative projects and findings, shedding light on the transformative possibilities and inherent risks of Deepfake
technology. By examining the "light" and "shadow" of generative AI, we aim to foster a deeper understanding of its implications for society and inspire responsible
innovation in this rapidly evolving domain.
Here is his homepage.
The architecture and organization of cells in our bodies reveal information on our health status, and visual inspection of stained tissue samples has been the state-of-the-art to diagnose disease such as cancer for the past 150 years. Learning-based approaches for automated analysis of scanned tissue samples have recently made it all the way in to the clinic thanks to the power of deep convolutional neural networks. Successful automation of diagnostics depends on variability within the collected data, as well as consistency and quality of the data annotation, which is often done manually by expert pathologists. The most common stains used to reveal the morphological patterns of tissue is called Hematoxylin and Eosin, or H&E for short, and reveals tissue structure by staining cell nuclei and supportive tissue. Much more of the activities going on in a tissue sample can be revealed if also visualizing the different proteins in the tissue. Already when I started my PhD studies in computerized image processing in the late 1990s, we pioneered the elaboration with methods to increase the number of different proteins that can be visualized in the same tissue sample. Together with biochemists we developed an approach to repeatedly stain and wash away stain from tissue, so that multiple proteins could be targeted and quantified through image analysis. This approach has recently flourished, and ‘spatial proteomics’ was selected as ‘Method of the year’ by Nature in December 2024. We have also been closely involved in the development of image-based methods for reading the genetic code of mRNA in tissue, deciphering the messages controlling protein synthesis. This is called ‘spatial transcriptomics’, and was selected as ‘Method of the year’ by Nature in 2020. I will present our current work, on using digital image processing and learning based approaches to close the circle by combining advanced imaging-based methods to reveal both function and morphology in pathology. Here is her homepage.
| Schedule | |||
|---|---|---|---|
| Wednesday, March 12 | Location | ||
| 9.45 - 10.00 | Student Registration | RPL | |
| 10.00 - 12.00 | PhD students morning | RPL | |
| 11.30 - 17.00 | Registration | Nymble | |
| 12.00 - 13.00 | Lunch | Nymble | |
| 13.15 - 13.30 | Welcome and opening | Nya Matsalen | |
| 13.30 - 14.30 | Keynote: Paolo Favaro Chair: Atsuto Maki | Advancing Autonomous Learning: Models that Learn, Adapt, and Predict from Real Data | Nya Matsalen |
| 14.30 - 14.55 | Oral presentation Chair: Atsuto Maki | • 14.30 - 14.55: Robust Camera Motion from Motion Blur | Nya Matsalen |
| 14.55 - 15.10 | Coffee break | Nymble | |
| 15.10 - 16.50 | Oral presentations Chair: Amanda Berg | • 15.10 - 15.35: TetraSphere • 15.35 - 16.00: Flopping for FLOPs Leveraging equivariance for computational efficiency • 16.00 - 16.25: SPL-BEV: Soccer Player Localization and Birds-Eye-View estimation • 16.25 - 16.50: Multimodal Analysis of Fragmentary Ancient Egyptian Papyri with Disentangled Features | Nya Matsalen |
| 17.00 - 18.00 | Reception | Secret location | |
| Thursday, March 13 | |||
| 9.00 - 10.40 | Industry presentations Chair: Mårten Wadenbäck | • 9.00 - 9.20: Univrses • 9.20 - 9.40: ContextVision • 9.40 - 10.00: Visual Sweden • 10.00 - 10.20: Savantic • 10.20 - 10.40: Maxar | Nya Matsalen |
| 10.40 - 10.55 | Coffee break | Nymble | |
| 10.55 - 12.10 | Oral presentations Chair: Pavlo Melnyk | • 10.55 - 11.20: Applying Center Loss to Neural Networks for Sequence Prediction: A Study for Handwriting Recognition • 11.20 - 11.45: A Study of Handwritten Text Recognition with Cross out Words • 11.45 - 12.10: Watch and Act - Multi-orientation Open-set Scene Text Recognition via Dynamic Expert Routing | Nya Matsalen |
| 12.10 - 13.10 | Lunch | Nymble | |
| 13.10 - 14.10 | Keynote: Toshihiko Yamasaki Chair: Atsuto Maki | Deepfake: The Light and Shadow of Generative AIs | Nya Matsalen |
| 14.10 - 15.00 | Oral presentations Chair: Matteo Gamba | • 14.10 - 14.35: Comparing Satellite Data for Next-Day Wildfire • 14.35 - 15.00: Energy-guided Decoding for Object Hallucination Mitigation | Nya Matsalen |
| 15.00 - 15.15 | Coffee break | Nymble | |
| 15.15 - 15.40 | Oral presentation Chair: Matteo Gamba | • 15.15 - 15.40: TAG: Text Prompt Augmentation for Zero-Shot Out-of-Distribution Detection | Nya Matsalen |
| 15.40 - 16.00 | Poster pitches Chair: Mårten Björkman | Nya Matsalen | |
| 16.00 - 17.00 | Poster session | • Physics-informed deep learning for predicting the arterial input function in dynamic PET imaging • The Impact of Data Pre-Processing on Multi-Modal Deep Learning in Spatial Transcriptomics • Handwritten Text Recognition for Historical Writings with Rare and Unknown Scripts: The DESCRYPT Project • Individually Fair Representation Learning for DINOv2 • Towards Controllable Image Generation through Representation-Conditioned Diffusion Models | Hyllan |
| 17.00 - 18.00 | SSBA annual meeting | Nya Matsalen | |
| 19.00 - | Dinner | Syster O Bror | |
| Friday, March 14 | |||
| 9.00 - 10.40 | Oral presentations Chair: Victor Wåhlstrand | • 9.00 - 9.25: Semantic Labeling of Persons in Point Clouds for Long Range Applications • 9.25 - 9.50: Reconstruction of Ultrasound-Speed Maps with a Learned Imaging Model • 9.50 - 10.15: Global Tissue Speed-of-Sound via Comparative and Absolute Image Metrics in Ultrasound • 10.15 - 10.40: Towards Out-of-Distribution Detection for Breast Cancer Classification in Point-of-Care Ultrasound Imaging | Nya Matsalen |
| 10.40 - 10.55 | Coffee break | Nymble | |
| 10.55 - 11.55 | Keynote: Carolina Wählby Chair: Mårten Björkman | Functional pathology – the power of combining image analysis and biochemistry to understand disease | Nya Matsalen |
| 11.55 - 12.20 | Oral presentation Chair: Mårten Björkman | • 11.55 - 12.20: Find ’em all Evaluating Deep Learning-based Cell Detection Methods | Nya Matsalen |
| 12.20 - 13.20 | Lunch | Nymble | |
| 13.20 - 14.35 | Oral presentations Chair: Elisa Barney Smith | • 13.20 - 13.45: Early Fusion of H&E and IHC Histology Images for Pediatric Brain Tumor Classification • 13.45 - 14.10: Explainable vertebral fracture analysis with uncertainty estimation using differentiable rule-based classification • 14.10 - 14.35: Bacterial species identification using deep learning-based image- and video classification of phase-contrast microscopy time-lapse videos | Nya Matsalen |
| 14.35 - 14.40 | Closing | Nya Matsalen | |
| Student registration without dinner (deadline February 28) | Regular registration (deadline February 28) |
|---|---|
| 2400 SEK | 2900 SEK |
