About MHaulBnech

[Estrellajer]

Thank you for your excellent work and for sharing this project.

I noticed that the dataset currently hosted on Hugging Face does not seem to contain the full set of images. According to the paper, the data was sampled from sources like MS COCO, which can be quite complex to download and reassemble manually.

Could you please provide a clearer guide or script on how to download the full image set and properly load the dataset for training/evaluation?

Thanks again for your help!

[sunnychenxiwang]

Thank you for your kind words and for raising this important question!

You're absolutely right: the version we uploaded to Hugging Face primarily contains the annotation metadata. Due to copyright restrictions and storage limitations, we are unable to redistribute the original images. Users need to download the images from their respective source datasets themselves.

The dataset was assembled by a team of contributors from multiple publicly available sources. Currently, we do not have a unified one-click download script, as the images come from different origins with their own download procedures and licensing terms.

For complete guidance on obtaining the full image set and correctly loading the dataset for training or evaluation, we recommend referring to the official repositories or documentation of the corresponding source datasets.

Thanks again for your interest in EasyDetect!

[Estrellajer]

Thank you for the quick response! I completely understand why the full images cannot be hosted directly.

To clarify, my main concern is the current workflow for obtaining the images. Right now, it seems the only way to identify the required images is by matching the image IDs in the JSON file.

However, downloading the entire source datasets (e.g., TextVQA, COCO) just to extract the specific subset used in your benchmark is quite a heavy task for many users.

To make your work more accessible, would it be possible to provide:

1.A downloader script: A Python script that takes the IDs from your JSON files and automatically fetches only the required images from the original official sources?

2.A clear mapping file: A simple text/CSV file listing exactly which Image IDs or file names from the source datasets (e.g., 'train2014/COCO_train2014_000000.jpg') are included in your benchmark?

This would greatly help users reproduce your results without having to manually filter or download tens of gigabytes of unnecessary data.

Thank you for your time and for the great contribution!