Corss Task Transferability - python scripts about sga HOT 3 CLOSED

I encountered the same problem, did you solve it?

from sga.

Hi, thanks for your interest in our work.
Here are the details regarding cross-task transferability in our paper:

1. The related experiments are conducted using the MS-COCO dataset, with ALBEF as the source model;
2. For crafting adversarial example, please refer to SGA repo. (https://github.com/Zoky-2020/SGA), specifically './SGA/eval_albef2tcl_flickr.py';
3. For the evaluation of visual grounding, refer to the ALBEF repo. (https://github.com/salesforce/ALBEF/tree/main), particularly './ALBEF/Grounding.py';
4. For the evaluation of image captioning, refer to the BLIP repo. (https://github.com/salesforce/BLIP/tree/main), particularly './BLIP/train_caption.py';

from sga.

Thanks for the response.

I have reproduced results for Table 4 (Image Caption - IC). I used ALBEF model and generated adversarial images from test set of MSCOCO. I later evaluated BLIP model on the adversarial test images.

Could you please clarify which set (train or test of MSCOCO) you used for Table 5 (Visual Grounding - VG)
It appears that visual grounding evaluation code of ALBEF is using train images.

from sga.