VideoNavQA

VideoNavQA: Bridging the Gap between Visual and Embodied Question Answering
BMVC 2019, spotlight talk at ViGIL NeurIPS 2019
Cătălina Cangea, Eugene Belilovsky, Pietro Liò, Aaron Courville

We introduce the VideoNavQA task: by removing the navigation and action selection requirements from Embodied QA, we increase the difficulty of the visual reasoning component via a much larger question space, tackling the sort of complex reasoning questions that make QA tasks challenging. By designing and evaluating several VQA-style models on the dataset, we establish a novel way of evaluating EQA feasibility given existing methods, while highlighting the difficulty of the problem even in the most ideal setting.

'Where is the green rug next to the sofa?'	'Are the computer and the bed the same color?'	'What is the thing next to the tv stand located in the living room?'

Getting started

$ git clone https://github.com/catalina17/VideoNavQA
$ virtualenv -p python3 videonavqa
$ source videonavqa/bin/activate
$ pip install -r requirements.txt

Dataset

The VideoNavQA benchmark data can be found here. After expanding the archive to a specific directory, please update BASE_DIR (declared in eval/utils.py) with that path.

Dependencies

• Model evaluation:
  • Faster-RCNN fork (with VGG-16 pre-trained weights)
  • pre-trained object detector for extracting visual features (OBJ_DETECTOR_PATH in eval/utils.py) should be initialised from this checkpoint instead of the one initially provided in the dataset archive - please make sure to replace the file!

• Data generation tools:
  • EmbodiedQA fork
  • House3D fork
  • SUNCG dataset
  • SUNCG toolbox

Running the models

The sample script eval.sh allows running (as-is) the FiLM-based models described in our paper. One epoch takes a few hours on an Nvidia P100 16GB GPU; it is likely that you will need to resume training from the specified checkpoint every 1-3 epochs. You may then test your model using the q_and_v_test.py script, with similar command-line arguments.

Citation

Please cite us if our work inspires your research or you use our code and/or the VideoNavQA benchmark:

@article{cangea2019videonavqa,
  title={VideoNavQA: Bridging the Gap between Visual and Embodied Question Answering},
  author={Cangea, C{\u{a}}t{\u{a}}lina and Belilovsky, Eugene and Li{\`o}, Pietro and Courville, Aaron},
  journal={arXiv preprint arXiv:1908.04950},
  year={2019}
}