This fork of bottom-up-attention is used to support feature extraction of faster-rcnn, which can be used through the pre-built Docker image chenrocks/butd-caffe:nlvr2.

Running Docker

bash scripts/extract_imgfeat.sh $PATH_TO_IMG_FOLDER $PATH_TO_IMG_NPY

Extracting Feature

Modify line 42-43 in 'tools/generate_npz.py' as follows.
line 42 MIN_BOXES = 100 -> MIN_BOXES = 36
line 43 MAX_BOXES = 100 -> MAX_BOXES = 36

```
python tools/generate_npz.py --def models/vg/ResNet-101/faster_rcnn_end2end_final/test.prototxt --cfg experiments/cfgs/faster_rcnn_end2end_resnet.yml --out /output/ --net 
data/faster_rcnn_models/resnet101_faster_rcnn_final.caffemodel --prefix butd

```
The output files will be stored in $PATH_TO_IMG_NPY as follows.
sample.jpg -> butd_sample.npz
(i.e. imgname.jpg -> butd_imgname.npz)