I2ID: Disentangling Identity Features via Synchronized Masking for Zero-Shot Composed Person Retrieval
(Accepted by Pattern Recognition, 2026)

Code

The source code is being organized.

📊 TG-CPR Dataset Overview

The statistics are as follows:

Dataset	Query IDs	Gallery IDs	Queries	Gallery Images	Public	Download	Paper
DeepChange	409	1,098	1,180	2,301	Per Request	link	link
CCVID	8	226	31	350	Yes	link	link
COCAS-Real2	87	95	174	190	Yes	link	link
VCCR	102	383	273	703	Yes	link	link
Total	606	1,802	1,658	3,544	-	-	-

📁 JSON File Structure

gallery.json

Gallery set containing candidate images for retrieval.

[
  {
    "file_path": "ccvid/session3/001_03/00252.jpg",
    "datasets": "ccvid",
    "person_id": 1,
    "instance_id": 1
  }
]

Field	Type	Description
file_path	string	Relative path to the image file
datasets	string	Source dataset name (deepchange, ccvid, cocas_real2, VCCR)
person_id	int	Unique identifier for each person identity
instance_id	int	Unique identifier for each person instance across datasets

query.json

Query set for standard person re-identification evaluation.

[
  {
    "file_path": "ccvid/session1/001_01/00065.jpg",
    "datasets": "ccvid",
    "person_id": 1,
    "instance_id": 1,
    "caption": "wearing a white polo shirt and dark pants"
  }
]

Field	Type	Description
file_path	string	Relative path to the query image
datasets	string	Source dataset name
person_id	int	Unique identifier for each person identity
instance_id	int	Unique identifier for each person instance
caption	string	Textual description of the person's appearance

query_and_gallery.json

[
  {
    "file_path": "ccvid/session1/001_01/00065.jpg",
    "datasets": "ccvid",
    "person_id": 1,
    "instance_id": 1,
    "caption": "wearing a white polo shirt and dark pants",
    "answer": "ccvid/session3/001_03/00252.jpg"
  }
]

Field	Type	Description
file_path	string	Relative path to the query image
datasets	string	Source dataset name
person_id	int	Unique identifier for each person identity
instance_id	int	Unique identifier for each person instance
caption	string	Textual description of the person's appearance
answer	string	Relative path to the ground-truth gallery image

📝 License

Each sub-dataset has its own license terms. Please refer to the original dataset sources for specific usage terms.

Important:

• All datasets are for non-commercial research purposes only
• Users must follow the original datasets' terms and conditions
• Face blocking may be required when publishing images

🙏 Acknowledgements

We thank the authors of the following datasets:

• DeepChange: Peng Xu, Xiatian Zhu
• CCVID: Xinqian Gu, Hong Chang, et al.
• COCAS-Real2: Shihua Li, Haobin Chen, et al.
• VCCR: ke Han, Yan Huang, et al.

📧 Contact

For questions about this aggregated dataset, please open an issue in this repository.