Update (29 Sep, 2021):

1. We realize that we've made changes to the source code of detectron2 repo in order to run feature extraction described in LuoweiZhou/VLP under detectron2 framework (because detectron1 is almost deprecated). For reproducibility, we release the source code under WebQA_x101fpn/detectron2/ in this repo. There is no need to install the detectron2 package from facebookresearch/detectron2!

2. We clarify here what are arguments --gold_feature_folder, --distractor_feature_folder, and --x_distractor_feature_folder. Basically, during implementation we divide the images into 3 buckets: positive images for image-based queries (gold), negative images for image-based queries (distractors) and negative images for text-based queries (x_distractors), where the 'x' stands for 'cross-modality'. Image- and text-based queries can be disinguished via the "Qcate" field in the dataset file. Text-based queries all have Qcate == 'text', while the rest are image-based ones.