Disclaimer: 99% of job is done by airsplay and adrelino.
In this repo, you can extract the features from Faster RCNN trained on Visual Genome provided originally by UpDown. This model is used in many papers, mostly because it is trained to predict not only bounding box but also its attributes.
In this repo, we extend the adrelino 's repo to FasterRCNN with attributes.
detectron2
torch >= 1.4.0
opencv-python
tqdm
It is highly recommended that you create a virtual environment. I have used miniconda3 with python3.7.7. Create the environment from the env.yml file:
conda env create --name <env_name> -f env.yml
You will not have detectron inside env.yml. To obtain the same detectron2 I have, please run:
python -m pip install 'git+https://github.com/facebookresearch/detectron2.git@ffff8ac'
If that fails as well, Google is your best friend.
Pretrained models thankfully converted from Caffe to Pytorch by airsplay
More info can be found here.
You can use the command line and its options:
python extract.py --image_dir ...
More information is coming...