This repository was cloned from an earlier version of faster-rcnn.pytorch repository.
It contains scripts to extract FasterRCNN features for CLEVR dataset.
You can download extracted features from here.
It has two keys:
image_featuresof shapeN x num_objects x feature_dims, contains the image features of 15 objects per image of 2048 dimensions.box_featuresof shapeN x num_objects x 6, where the 6 dimensions refer to: (x1, y1, x2, y2, width, height). They are normalized to 0 and 1.
-
Compile the library by executing
make.shinside thelibdirectory. Note that I had faced several issues while compiling the library. I used the following setup/modifications, which may be helpful to you too:a. It requires Pytorch version: 0.4.0 (Versions 0.4.1 and 1.0 do not work!). You can install the correct dependencies using:
conda install pytorch=0.4.0 torchvision -c pytorchb. You may have to edit the
CUDA_ARCHvariable insidelib/make.shto ensure things are compatible with your GPU. -
Download pre-trained FasterRCNN model to a path, say, to:
${ROOT}/FasterRCNN/models/res101/clevrThis model has been trained on training images of CLEVR dataset. -
Download objects_count.json inside
${ROOT}/CLEVR/faster-rcnn/ -
Put CLEVR images inside the following directories:
a. Train images inside
${ROOT}/CLEVR/images/trainb. Val images inside
${ROOT}/CLEVR/images/valc. Test images inside
${ROOT}/CLEVR/images/test -
Execute
./extract_resnet_features_CLEVR.shThis will extract the features to${ROOT}/CLEVR/features
Here is the link to the original repository.