Yash Kant, Dhruv Batra, Peter Anderson, Alex Schwing, Devi Parikh, Jiasen Lu, Harsh Agrawal
Published at ECCV, 2020
Paper: arxiv.org/abs/2007.12146
Project Page: yashkant.github.io/projects/sam-textvqa
We propose a novel spatially aware self-attention layer such that each visual entity only looks at neighboring entities defined by a spatial graph and use it to solve TextVQA.
Create a fresh conda environment, and install all dependencies.
conda create -n sam python=3.6
conda activate sam
cd sam-textvqa
pip install -r requirements.txt
Install pytorch
conda install pytorch torchvision cudatoolkit=10.0 -c pytorch
Finally, install apex from: https://github.com/NVIDIA/apex
Download files from the dropbox link and place it in the data/ folder.
Ensure that data paths match the directory structure provided in data/README.md
From the below table pick the suitable configuration file:
| Method | context (c) | Train splits | Evaluation Splits | Config File |
|---|---|---|---|---|
| SA-M4C | 3 | TextVQA | TextVQA | train-tvqa-eval-tvqa-c3.yml |
| SA-M4C | 3 | TextVQA + STVQA | TextVQA | train-tvqa_stvqa-eval-tvqa-c3.yml |
| SA-M4C | 3 | STVQA | STVQA | train-stvqa-eval-stvqa-c3.yml |
| SA-M4C | 5 | TextVQA | TextVQA | train-tvqa-eval-tvqa-c5.yml |
To run the experiments use:
python train.py \
--config config.yml \
--tag experiment-name
To evaluate the pretrained checkpoint provided use:
python train.py \
--config configs/train-tvqa_stvqa-eval-tvqa-c3.yml \
--pretrained_eval data/pretrained-models/best_model.tar
Note: The beam-search evaluation is undergoing changes and will be updated.
Resources Used: We ran all the experiments on 2 Titan Xp gpus.
@inproceedings{kant2020spatially,
title={Spatially Aware Multimodal Transformers for TextVQA},
author={Kant, Yash and Batra, Dhruv and Anderson, Peter
and Schwing, Alexander and Parikh, Devi and Lu, Jiasen
and Agrawal, Harsh},
booktitle={ECCV}
year={2020}}
Parts of this codebase were borrowed from the following repositories:
- 12-in-1: Multi-Task Vision and Language Representation Learning: Training Setup
- MMF: A multimodal framework for vision and language research: Dataset processors and M4C model
We thank Abhishek Das, Abhinav Moudgil for their feedback and Ronghang Hu for sharing an early version of his work. The Georgia Tech effort was supported in part by NSF, AFRL, DARPA, ONR YIPs, ARO PECASE, Amazon. The views and conclusions contained herein are those of the authors and should not be interpreted as necessarily representing the official policies or endorsements, either expressed or implied, of the U.S. Government, or any sponsor.
MIT