A pre-generated version of the dataset as used in the paper is provided here.
There are 4 folders at the link, corresponding to the different versions of the data used for experiments in the paper. Each contains a zip file with the videos, metadata, and a folder with the train/val files. The train files are further split into train_subsetT.txt (80% of train.txt) and train_subsetV.txt (20% of train.txt) as validation sets that could be used for hyperparameter optimization.
max2action: With only 2 objects moving in each time segment. We use these for evaluating the atomic and compositional recognition tasks (Tasks 1 and 2).all_actions: With all objects moving at each time segment. We use these for evaluating the localization task (Task 3).
For both, a version with camera motion is also provided. Each also contains a list folder with subfolder for each task setup (localize, actions_order_uniq for compositional, and actions_present for atomic). The localize task also has a 4x4 and 8x8 setting which were used for ablating the grid sizes. Each of those subfolders have a train.txt and val.txt files, which are formatted as video_id, class(es).
Here are the direct links to the large ZIP files, which should make it easier to download on headless servers.
- all_actions
- all_actions_cameramotion
- max2action
- max2action_cameramotion
This data is also available on baidu (code: d9nk) for faster access in Mainland China, thanks to ruiyan1995.
If you'd like to regenerate CATER or generate it with some modifications, please read on.
- All CLEVR requirements (eg, Blender: the code was tested with v2.79b).
- This code was tested on CentOS 7. However, since not all dependencies were available locally in my machine, I ran it using in a singularity VM. Here's the VM spec that was used in the following code. If you need it, download and set the path in
launch.py. - GPU: This code was tested with TITAN-X GPUs, though it should be compatible with most NVIDIA GPUs. By default it will use all the GPUs on the machine.
Run python launch.py to start generating. Please read through the launch script to change any settings, paths etc. The command line options should also be easy to follow from the script. If using singularity, you'll need to set a data mount dir, and store videos w.r.t that path.
You can use the gen_train_test.py script to generate labels for the dataset for each of the tasks. Change the parameters on the top of the file, and run it.