Looking using bilinear features for classification.
This code (approximately) reproduces the results of
http://vis-www.cs.umass.edu/bcnn/docs/bcnn_iccv15.pdf
Need to download the CUB_200_2011 dataset and rename ./data/cub
There are a number of python dependencies -- need to get around to adding a requirements.txt. In the meantime, the only unusual dependency is:
- tdesc -- https://github.com/bkj/tdesc
INPATH=./data/cub/images/
# Regular features
find $INPATH -type f | python -m tdesc --model vgg16 --crow --target-dim 448 > ./data/feats-crow-448
# Get last convolutional features, then expand to bilinear features
find $INPATH -type f | ./get_conv.py --target-dim 448 --outpath ./conv.bc > ./data/convpaths
./bilinear.py --inpath ./conv.bc --outpath ./data/bilinear.bc
# run ./billy.py