GloViTa is a reusable training framework for supervised classification and regression on image data, developed by the Applied Computer Vision Lab at DKFZ. It is domain-agnostic: the same code applies to microscopy, satellite imagery, medical scans, industrial inspection images, or pathology slides, provided the task can be expressed as predicting a label or a continuous value from an image. Supported input types include 2D images, 3D volumes, video, and bag-of-patches representations such as whole-slide images.

The framework replaces per-project training scripts with a configuration-driven workflow. A run is specified by a small number of options — the data location, the number of target classes, the pretrained backbone — and launched with a single command. Data loading, training, and validation are handled by the framework, and the full configuration of each run is written to disk, which keeps experiments reproducible and comparable.

The same design makes the framework useful for systematic experimentation. Because all settings are exposed through one interface, varying the learning rate, augmentation strength, or model size requires changing an option rather than modifying code. Several regularization and finetuning strategies are included, ranging from parameter-efficient adaptation of a pretrained model to full finetuning, so comparing them on a given dataset does not require implementing them first.

The intended users are researchers and practitioners who are comfortable with command-line tools, and who would prefer not to maintain separate training code for each new dataset. For groups where imaging tasks recur across projects, this removes a substantial amount of duplicated engineering effort.

Code: github.com/MIC-DKFZ/image_classification

Developed within the Applied Computer Vision Lab, the Helmholtz Imaging Support Unit at DKFZ. Apache-2.0 licensed.