How to generate my own vgbua_res101_precomputed files.

[kememem]

Hi,

When I use your code to evaluate my own visual grounding dataset, could you give me some instructions for constructing vgbua_res101_precomputed files?

Thanks

[jackroos]

You can download the files for VQA here and follow the same format. Thank you!

[kememem]

Thanks for the reply. I've checked the files and understand the meaning of most key/value pairs. However, could you explain me what does "boxes" key mean? How can I generate the values of "boxes" from an annotated image. Thanks!

[jackroos]

It means the bounding boxes of objects in the image, you can decode the "boxes" following our code of VQA dataset:

VL-BERT/vqa/data/datasets/vqa.py

Line 445 in d58c70c

item[field] = np.frombuffer(base64.decodebytes(item[field].encode()),

After it is decoded, it should be a numpy array of shape (#boxes, 4). And you can encode boxes of your own image by following way:
https://github.com/jackroos/bottom-up-attention/blob/4de0a37c98039470f27a0cbfa94765e90bb80d24/tools/generate_tsv_v2.py#L98