Extracting image features
Edit config settings in
config.py, then run
Tip: If you have multiple GPUs, you can parallelize the feature extraction by extracting
test2015 on separate GPUs. You’ll need to change the code in
Preparing the Visual Genome QA pairs
Clone the Visual Genome Python Driver repo into this directory. Download images, image meta data, and question answers from Visual Genome website. The image meta data json file and question answers json file should be in a directory named genome. Then use
vg_preprocessing.py to convert the Visual Genome QA file(s) into the format we need.