TPT does not explicitly align the pre-trained CLIP to become aware of the test sample distribution. For the effective test-time adaptation of V -L foundation models, it is crucial to bridge the distribution gap between the pre-training dataset and the downstream evaluation set for high zero-shot generalization.

Evaluating machine learning models requires evaluation data.

Competing objectives arise when a model's capabilities

We demonstrate FlexCap's effectiveness in several applications: first, it achieves strong performance in dense captioning tasks on the Visual Genome dataset. Second, we show how FlexCap's localized

The croissant metadata record is available at croissant. We chose GitHub and Google Drive respectively to store our code and dataset. Both are widely recognized as reliable data storage platforms, ensuring long-term preservation. We highly recommend downloading the raw data directly and following the provided instructions to simplify the data processing steps. Our dataset is structured as follows: the local directory contains client-specific data for local training, while all clients aggregates data from all clients for federated learning.