adobe-research / vaw_dataset Goto Github PK

Hi @kushalkafle, thanks for your great works of VAW and LSA. And I have some questions about the inference details of the SCoNE and TAP.
During inference，
For SCoNE, did you crop out the object region first and then evaluate the precision of the method for each bounding box?
For TAP and OpenTAP, did you just input the test image and multi objects with bounding boxes, then the model will output the attributes of each object?
I wonder if the above conjectures match the real experimental design. Looking forward to your reply and thanks in advance!

Hi, I am a little confused about the object embedding procedure. As mentioned in the paper, GloVe 100-d word embeddings are used as the object name embedding. However, some of the object names are not contained in the Glove embeddings. How to tackle these names? For example, 'american flag', "boy's arm", 'two suitcases', 'computer keyboard', 'larger horse', 'living room wall', 'navy blue shirt', 'of the aisle', 'hotdog bun', 'train station', 'skull picture', 'disney princess', 'neck tie'.

Thanks.

Thank you for the great work and code!
Do you have any plans on uploading code for inference?
Thank you

I counted the number of child attributes in attribute_type.json but there are 634 attribute after removing the duplicates. Why does README say 620 unique attributes?

Thanks for releasing the VAW dataset. I wonder whether the code of the method proposed in the paper will also be publicly available?

Hi,
Thank you for releasing the attribute annotations. A am very interested in the dataset.
Are you also planning to release the superclass list of attributes from the paper (the Class imbalance and Attribute types)?
And could you provide your evaluation code to reproduce your results and use the dataset?

Best,
Maria