Idan/visdial-fga-preprocessed

VisDial v1.0, preprocessed for Factor Graph Attention

The preprocessed VisDial v1.0 files used by Factor Graph Attention (CVPR'19) — code at idansc/fga.

Evaluation is done on VisDialv1.0.

Short description:

VisDial v1.0 contains 1 dialog with 10 question-answer pairs (starting from an image caption) on ~130k images from COCO-trainval and Flickr, totalling ~1.3 million question-answer pairs.

These are the tokenized, integer-indexed versions of those dialogs: every question, answer, caption and candidate answer is stored as vocabulary ids, which is the form the model consumes. They are published here because the links previously given in the repository have expired.

Files

File	Contents
visdial_data.h5	Tokenized dialogs, answer options and captions for train/val/test
visdial_params.json	Vocabulary (word2ind, ind2word) and the image filename list per split
visdial_1.0_val_dense_annotations.json	Dense relevance judgements for the val split, used for NDCG

visdial_data.h5 layout

Per split (train / val / test):

Dataset	Shape	Meaning
ques_{split}	(num_images, 10, 20)	Question token ids, zero-padded after the last word
ques_length_{split}	(num_images, 10)	Words per question
ans_{split}	(num_images, 10, 20)	Ground-truth answer token ids
ans_length_{split}	(num_images, 10)	Words per answer
opt_{split}	(num_images, 10, 100)	1-based indices into opt_list_{split}
opt_list_{split}	(num_answers, 20)	The pool of unique candidate answers
opt_length_{split}	(num_answers,)	Words per candidate answer
ans_index_{split}	(num_images, 10)	1-based index of the correct option (train/val only)
cap_{split}	(num_images, 40)	Caption token ids
cap_length_{split}	(num_images,)	Words per caption
num_rounds_{split}	(num_images,)	Rounds actually present; the test split is cut at a random round
img_pos_{split}	(num_images,)	Image position index

Splits are 123,287 / 2,064 / 8,000 images. The vocabulary holds 11,319 words; token id 0 is padding, and the loader appends a <stop> and an <empty> symbol, so an embedding table needs 11,322 rows.

Image filenames live in visdial_params.json under unique_img_{split}, e.g. VisualDialog_val2018/VisualDialog_val2018_000000185565.jpg.

Usage

from huggingface_hub import snapshot_download
from fga import VisDialDataset
from fga.data import load_visdial_params, vocab_size_from_params

path = snapshot_download("Idan/visdial-fga-preprocessed", repo_type="dataset")
params = load_visdial_params(f"{path}/visdial_params.json")

dataset = VisDialDataset(
    visdial_data_path=f"{path}/visdial_data.h5",
    image_features_path="frcnn_features_new.h5",   # see below
    split="val",
    vocab_size=vocab_size_from_params(params),
)

What is not here

Image features. The model reads pre-extracted region features, not pixels — an h5 with {split}_features of shape (num_images, 37, 2048). Those are not included here.

Pretrained features:

• VGG: a grid image feature based on the VGG model pretrained on ImageNet (Faster). Note, the h5 databases has slightly different dataset keys, therefore the code needs to be adapted accordingly.
• F-RCNN: based on object detector with ResNeXt-101 backbone, 37 proposals, fine-tuned on Visual Genome. Achieves SOTA. The file includes boxes and classes information.

See the original paper for performance differences. I recommend using the FRCNN features, mainly because it is finetuned on the relevant VisualGenome dataset.

The images themselves can be obtained from visualdialog.org/data or from the HuggingFaceM4/VisDial mirror. Note that the Hub mirrors of VisDial carry only [question, answer] pairs — they do not include the answer options or the ground-truth index, so they cannot be used to train or evaluate a ranking model on their own.

Licensing and credit

Released under CC BY 4.0, matching the upstream VisDial v1.0 annotations. This is a preprocessed derivative; the underlying dialog data is by Das et al., and the images come from COCO and Flickr under their own terms.

@inproceedings{das2017visual,
  title={Visual Dialog},
  author={Das, Abhishek and Kottur, Satwik and Gupta, Khushi and Singh, Avi and Yadav, Deshraj
          and Moura, Jos\'e M.F. and Parikh, Devi and Batra, Dhruv},
  booktitle={Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition},
  year={2017}
}

Please cite Factor Graph Attention if you use this work in your research:

@inproceedings{schwartz2019factor,
  title={Factor graph attention},
  author={Schwartz, Idan and Yu, Seunghak and Hazan, Tamir and Schwing, Alexander G},
  booktitle={Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition},
  pages={2039--2048},
  year={2019}
}