Word4Per is an innovative framework for Zero-Shot Composed Person Retrieval (ZS-CPR), integrating visual and textual information for enhancing person identification. This repository includes the Word4Per code and the Image-Text Composed Person Retrieval (ITCPR) dataset, offering new tools for research in security and social applications.

• [2023.11.16] Repo is created. Code and Dataset will come soon.
• [2023.11.25] The ITCPR dataset is now publicly available for download.

The ITCPR dataset is a comprehensive collection specifically designed for the Zero-Shot Composed Person Retrieval (ZS-CPR) task. It consists of a total of 2,225 annotated triplets, derived from three distinct datasets: Celeb-reID, PRCC, and LAST. To access the ITCPR dataset, please use the following download link: ITCPR Dataset Download.

• Total Annotated Triplets: 2,225
• Unique Query Combinations: 2,202
• Total Images: 1,151 from Celeb-reID, 146 from PRCC, 905 from LAST
• Total Identities: 512 from Celeb-reID, 146 from PRCC, 541 from LAST
• Target Gallery: 20,510 images with 2,225 corresponding ground truths

The images in the ITCPR dataset are sourced from the following datasets:

• Celeb-reID
• PRCC
• LAST

These are utilized solely for testing purposes in the ZS-CPR task.

The dataset includes two annotation files: query.json and gallery.json.

Each entry in the query.json file follows this structure:

{
    "file_path": "Celeb-reID/001/1_1_0.jpg",
    "datasets": "Celeb-reID",
    "person_id": 1,
    "instance_id": 1,
    "caption": "Wearing a brown plaid shirt, black leather shoes, another dark gray T-shirt, another blue jeans"
}

• file_path: Reference image path relative to the data root directory.
• datasets: Source dataset of the image.
• person_id: Person ID in the original dataset.
• instance_id: Unique identifier for gallery ground truth matching.
• caption: Relative caption of the reference image.

Each entry in the gallery.json file follows this structure:

{
    "file_path": "Celeb-reID/001/1_2_0.jpg",
    "datasets": "Celeb-reID",
    "person_id": 1,
    "instance_id": 1
}

• instance_id: Matches with query.json for target images; -1 for non-matching query instances.
• Others: Correspond to target image path, original dataset, and person ID.

data
|-- Celeb-reID
|   |-- 001
|   |-- 002
|   |-- 003
|   ...
|-- PRCC
|   |-- train
|   |-- val
|   |-- test
|-- LAST
|   |-- 000000
|   |-- 000001
|   |-- 000002
|   ...
|-- query.json
|-- gallery.json

Download and prepare the datasets as follows:

1. Celeb-reID: GitHub Repository
2. PRCC: Google Drive Link
3. LAST: GitHub Repository

After downloading, use the img_process.py script to process Celeb-reID and LAST datasets into the standard format. The PRCC (subfolder PRCC/rgb) dataset can be directly placed in the corresponding directory upon extraction.

We are deeply thankful to the creators of the Celebrities-ReID, PRCC, and LAST datasets for their significant contributions to the field of person re-identification. Their commitment to open-sourcing these valuable resources has greatly facilitated advancements in academic and practical research.

• Celebrities-ReID: "Celebrities-ReID: A Benchmark for Clothes Variation in Long-Term Person Re-Identification" - View Paper
• PRCC: "Person Re-identification by Contour Sketch under Moderate Clothing Change" - View Paper
• LAST: "Large-Scale Spatio-Temporal Person Re-identification: Algorithms and Benchmark" - View Paper

Certainly! You can use the following Markdown paragraph for your GitHub repository to instruct users to cite your paper if they utilize your code and dataset. Here's how you can format it:

If you use our code or dataset in your research, please cite our paper as follows:

@misc{liu2023word,
  title={Word for Person: Zero-shot Composed Person Retrieval},
  author={Delong Liu and Haiwen Li and Zhicheng Zhao and Fei Su and Hongying Meng},
  year={2023},
  eprint={2311.16515},
  archivePrefix={arXiv},
  primaryClass={cs.CV}
}