CountCLIP - [Re] Teaching CLIP to Count to Ten

Dataset We have created a small counting set of ~2000 images after passing over 2 million images out of the 400 million present in the original dataset. This is merged with ~13000 non-counting images from the same dataset. The entire merged dataset, along with the required relevant JSON/CSV files, can be found below. data.zip - merged counting and noncounting data, along with the validation data (the CountBench dataset). merged.json - JSON for merged (counting+noncounting) data. val.json - JSON for the CountBench data. faulty.csv - CSV for removing faulty noncounting images.