Official implementation of the Answer Mining from a Pool of Images: Towards Retrieval-Based Visual Question Answering (IJCAI 2023 paper)

paper | arxiv | project page

• Use python >= 3.8.5. Conda recommended : https://docs.anaconda.com/anaconda/install/linux/

• Use pytorch 1.9.0; CUDA 11.1

To setup environment

conda env create -n retvqa --file retvqa.yml
conda activate retvqa

Images: Visual Genome Krishna et al.

RetVQA: here

Image Feature extraction: Inside feature_extraction/, run

python retqa_proposal.py

Refer to this repo for more detailed set-up of Faster R-CNN feature extractor.

Refer to COFAR. ITM only variant serves as relevance encoder, when pre-trained on COCO and finetuned on RetVQA.

bash retvqa_VLBart.sh <no. of GPU>

bash retvqa_VLBart_test.sh <no. of GPU>

bash retvqa_retrieved_VLBart_test.sh <no. of GPU>

Download our retvqa finetuned checkpoint from here.

python eval_retvqa.py --gt_file <ground truth answers file path> --results_file <path to the generated answers>

This code and data are released under the MIT license.

If you find this data/code/paper useful for your research, please consider citing.

@inproceedings{retvqa,
  author       = {Abhirama Subramanyam Penamakuri and
                  Manish Gupta and
                  Mithun Das Gupta and
                  Anand Mishra},
  title        = {Answer Mining from a Pool of Images: Towards Retrieval-Based Visual
                  Question Answering},
  booktitle    = {IJCAI},
  publisher    = {ijcai.org},
  year         = {2023},
  url          = {https://doi.org/10.24963/ijcai.2023/146},
  doi          = {10.24963/ijcai.2023/146},
}

1. We used code-base and pre-trained models of VLBart.
2. Abhirama S. Penamakuri is supported by Prime Minister Research Fellowship (PMRF), Minsitry of Education, Government of India.
3. We thank Microsoft for supporting this work through the Microsoft Academic Partnership Grant (MAPG) 2021.