WISMIR3: A Multi-Modal Dataset to Challenge Text-Image Retrieval Approaches

This repository holds the WISMIR3 dataset. For more information, please refer to the paper:

@inproceedings{
schneider2024wismir,
title={{WISMIR}3: A Multi-Modal Dataset to Challenge Text-Image Retrieval Approaches},
author={Florian Schneider and Chris Biemann},
booktitle={3rd Workshop on Advances in Language and Vision Research (ALVR)},
year={2024},
url={https://openreview.net/forum?id=Q93yqpfECQ}
}

Download Images

To get the images:

1. Download all image parts from images directory.
2. Join the parts into a single archive file cat images/images.tar.gz.part* > images/images.tar.gz
3. Extract the archive tar xvzf images.tar.gz

Download Pre-computed Embeddings

To get the embeddings:

1. Download all image parts from the features directory. frcnn_feats are Faster-R-CNN Features, clip_ViT-B-16 are CLIP embeddings. For more information, please refer to the paper. 2.1. Join the parts into a single archive file cat features/clip_ViT-B-16_embs.tar.gz.part_* > features/clip_ViT-B-16_embs.tar.gz 2.2. Join the parts into a single archive file cat features/frcnn_feats.tar.gz.part_* > features/frcnn_feats.tar.gz 3.1. Extract the archive tar xvzf features/clip_ViT-B-16_embs.tar.gz 3.2. Extract the archive tar xvzf features/frcnn_feats.tar.gz

Columns

ColumnId	Description	Datatype
wikicaps_id	ID (line number) of the row in the original WikiCaps Dataset img_en	int
wikimedia_file	Wikimedia File ID of the Image associated with the Caption	str
caption	Caption of the Image	str
image_path	Local path to the (downloaded) image	str
num_tok	Number of Tokens in the caption	int
num_sent	Number of Sentences in the caption	int
min_sent_len	Minimum number of Tokens in the Sentences of the caption	int
max_sent_len	Maximum number of Tokens in the Sentences of the caption	int
num_ne	Number of Named Entities in the caption	int
num_nouns	Number of Tokens with NOUN POS Tag	int
num_propn	Number of Tokens with PROPN POS Tag	int
num_conj	Number of Tokens with CONJ POS Tag	int
num_verb	Number of Tokens with VERB POS Tag	int
num_sym	Number of Tokens with SYM POS Tag	int
num_num	Number of Tokens with NUM POS Tag	int
num_adp	Number of Tokens with ADP POS Tag	int
num_adj	Number of Tokens with ADJ POS Tag	int
ratio_ne_tok	Ratio of tokens associated with Named Entities vs all Tokens	int
ratio_noun_tok	Ratio of tokens tagged as NOUN vs all Tokens	int
ratio_propn_tok	Ratio of tokens tagged as PROPN vs all Tokens	int
ratio_all_noun_tok	Ratio of tokens tagged as PROPN or NOUN vs all Tokens	int
fk_re_score	Flesch-Kincaid Reading Ease score of the Caption ***	int
fk_gl_score	Flesch-Kincaid Grade Level score of the Caption ***	int
dc_score	Dale-Chall score of the Caption ***	int
ne_texts	Surface form of detected NamedEntities	List[str]
ne_types	Types of the detected NamedEntities (PER, LOC, GPE, etc.)	List[str]

---

See https://en.wikipedia.org/wiki/List_of_readability_tests_and_formulas for more information about Readability Scores

WikiCaps publication

WISMIR3 is based on the WikiCaps dataset. For more information about the WikiCaps, see https://www.cl.uni-heidelberg.de/statnlpgroup/wikicaps/

@inproceedings{schamoni-etal-2018-dataset,
    title = "A Dataset and Reranking Method for Multimodal {MT} of User-Generated Image Captions",
    author = "Schamoni, Shigehiko  and
      Hitschler, Julian  and
      Riezler, Stefan",
    editor = "Cherry, Colin  and
      Neubig, Graham",
    booktitle = "Proceedings of the 13th Conference of the Association for Machine Translation in the {A}mericas (Volume 1: Research Track)",
    month = mar,
    year = "2018",
    address = "Boston, MA",
    publisher = "Association for Machine Translation in the Americas",
    url = "https://aclanthology.org/W18-1814",
    pages = "140--153",
}