other
Hochschule der Medien Stuttgart
Institute for Applied Artificial Intelligence
We aim to develop novel architectures for deep learning based computer vision algorithms. The goal of the project is to develop architectures that are more robus to common corruptions and unusual images than existing convolutional neural networks or vision transformer. Specifically, we experiment with the Mask R-CNN framework which requires nodes with 8-GPUs (e.g. A100) per node to be trained on MS-COCO, a default benchmark dataset. On such nodes, one training takes approx. 12-24h. Without these compute reccources, iterative experimentation is not feasible in a reasonable time frame.