Poster
in
Workshop: ICML 2024 Workshop on Foundation Models in the Wild
Evaluating Self-Supervised Foundation Models in Holographic Imaging
Silas Dietler · Yanick Zeder · Elias Graf · Kilian Koch · Andreas Schwendimann · Tommaso Bendinelli
Keywords: [ DinoV2 ] [ Computer Vision ] [ Holographic Images ] [ Self-supervised training ]
DINOv2, a large self-supervised computer vision foundation model, has achieved impressive performance on downstream tasks like classification, segmentation, and depth estimation. This success suggests the idea that universal features can be extracted through large-scale pre-training. However, its applicability beyond natural image domains remains relatively unexplored. This study aims to contribute in this direction, by exploring the potential of DINOv2 for a niche but important task: pollen classification based on holographic images. Our findings reveal that features learned by the network in the natural image domain are not informative for this task. However, when DINOv2 is pre-trained on a pollen-specific dataset, it achieves superior performance compared to supervised methods, especially in scenarios with limited data. This superior performance opens doors for new applications such as online few-shot (bio)aerosol particle classification with holographic imaging.