Authors: Francesc Net, Lluis Gomez
Institution: Computer Vision Center, Universitat Autònoma de Barcelona
The EUFCC-CIR dataset is a cutting-edge resource designed to enable Composed Image Retrieval (CIR) within Galleries, Libraries, Archives, and Museums (GLAM) collections. It builds upon the EUFCC-340K image labeling dataset, providing over 180K meticulously annotated CIR triplets. Each triplet comprises:
- A query image 🖼️.
- A brief text description 📝 that specifies a desired modification.
- A target image 🎯 that fulfills the query.
This dataset is particularly beneficial for researchers in digital humanities, allowing them to retrieve and interact with cultural heritage items in new and engaging ways.
🔗 Paper link: EUFCC-CIR Paper
📂 Related dataset: EUFCC-340K
The dataset is organized within the data/ folder, consisting of the following key files:
db_processed.txt: 📄 A list of images, each assigned to a partition (train, validation, test).cir_db.csv: 🗃️ Contains detailed information about image-text pairs and their corresponding target images.
| 🔢 Column | 📖 Description |
|---|---|
| id1 | 🆔 Identifier for the first image in the pair. |
| id2 | 🆔 Identifier for the second image in the pair. |
| materials_1 | 🏗️ Materials in the first image. Specifies the substances or materials used in the first object. |
| ObjectTypes_1 | 🏷️ Object types in the first image. Defines the category or type of object represented. |
| materials_2 | 🏗️ Materials in the second image. Specifies the substances or materials used in the second object. |
| ObjectTypes_2 | 🏷️ Object types in the second image. Defines the category or type of object represented. |
| element_to_change | 🔄 The element (materials or object type) that must change between the two images. |
| element_changed | 🔄 The element (materials or object type) that was changed between the two images. |
| partition | 📊 Denotes the dataset split (train, validation, test) that this row belongs to. |
| query | 🔍 Describes the retrieval task, indicating how the transformation from id1 to id2 should occur. |
Note: Each row in the CSV represents a relationship between two images (
id1andid2), where an element in the first image is altered to generate the second image.
Below is an example visualization of how an image transformation is represented in this dataset:
To get started, clone the repository and download the necessary files from the EUFCC-340K dataset:
git clone https://github.com/your-username/EUFCC-CIR.git
cd EUFCC-CIR/data
# Download dataset files (from EUFCC-340K repo => link in the beginning)