MInterPEDES

MInterPEDES is a multi-turn multimodal conversational dataset for Multimodal Interactive Person Retrieval (MInterPR). It was introduced in the ICML 2026 paper “Interactive Person Retrieval via Multi-Turn Multimodal Conversation.”

MInterPR extends text-based and chat-based person retrieval by allowing a user to inspect retrieved reference images and provide corrective natural-language feedback about how each reference differs from the target person. MInterPEDES augments the question-answering dialogues of ChatPedes with reference-image paths and corrective feedback for studying this setting.

• Dataset repository: byougert/MInterPEDES
• Official code: Flame-Chasers/MNEMO
• Original image source: CUHK-PEDES

Repository Contents

This repository releases annotations only:

MInterPEDES/
├── README.md
├── train_reid.jsonl
└── test_reid.jsonl

The original CUHK-PEDES images are not included or redistributed. The paths stored in the annotation files are relative paths into a separately obtained copy of CUHK-PEDES.

Dataset Statistics

Each target image is associated with two multimodal dialogue records constructed using two different reference-image sampling strategies. Consequently, the number of annotation records is twice the number of target images.

Split	Identities	Images	Dialogues
Train	11,003	34,054	68,108
Test	1,000	3,074	6,148
Total	12,003	37,128	74,256

The training and test splits are identity-disjoint and follow the CUHK-PEDES split protocol.

Data Format

Both files use JSON Lines format: each line is one JSON object representing one multimodal dialogue associated with a target image.

Field	Type	Description
file_path	string	The target image path.
id	integer	Person identity label.
t_conversation_length	integer	Number of text question-answer turns after the initial description.
i_conversation_length	integer	Number of visual feedback turns.
conversations	list of objects	Ordered dialogue messages. Every object contains from and value.
image	list of strings	Relative paths of the reference images, in the same order as the image turns in conversations.

The ordered conversations sequence has the following structure:

1. The first human message is the initial description of the target person.
2. The next 2 × t_conversation_length messages alternate between a gpt question and a human answer.
3. The final 2 × i_conversation_length messages alternate between a gpt image marker (<image>) and the corresponding human corrective feedback.

Thus, for each record:

len(conversations) = 1 + 2 * t_conversation_length + 2 * i_conversation_length
len(image) = i_conversation_length

The from labels follow the interaction convention used by the MNEMO data loader: the model asks a question or presents a retrieved image, and the user supplies an answer or corrective feedback.

Each file_path occurs exactly twice within its split. The released JSONL schema does not contain a separate field naming the sampling strategy; use the official MNEMO data loader when strategy-specific ordering is required.

Obtaining the Images

This repository does not grant access to, license, or redistribute the CUHK-PEDES images. To reconstruct the complete dataset:

1. Visit the official CUHK-PEDES repository.
2. Follow its current access procedure and terms of use to obtain the images. The upstream page states that access is for academic use and that requests should be made with a university email because of licensing issues.
3. Keep the directory names referenced by file_path and image, or adapt them through the data-preparation instructions in the MNEMO repository.

Users are solely responsible for complying with the CUHK-PEDES access conditions and any other applicable upstream terms.

Usage

Load from the Hugging Face Hub

pip install -U datasets

from datasets import load_dataset

dataset = load_dataset("byougert/MInterPEDES")
train_set = dataset["train"]
test_set = dataset["test"]

print(train_set[0])

The image-related fields remain relative path strings because the image files are not included.

Load JSONL directly

import json

with open("train_reid.jsonl", "r", encoding="utf-8") as f:
    sample = json.loads(next(f))

print(sample["file_path"])
print(sample["conversations"])
print(sample["image"])

Resolve image paths

After obtaining CUHK-PEDES, resolve target and reference images relative to the local image root:

from pathlib import Path
from datasets import load_dataset

sample = load_dataset("byougert/MInterPEDES", split="train")[0]

# This directory must directly contain the subdirectories referenced by
# file_path and image, such as CUHK01 or train_query.
image_root = Path("/path/to/CUHK-PEDES-image-root")

target_image = image_root / sample["file_path"]
reference_images = [image_root / path for path in sample["image"]]

For the authoritative training, evaluation, and directory-preparation workflow, use the official MNEMO code.

License

The original annotation content released in train_reid.jsonl and test_reid.jsonl is licensed under the Creative Commons Attribution-NonCommercial 4.0 International License (CC BY-NC 4.0).

This license applies to the MInterPEDES annotations only. It does not apply to CUHK-PEDES images or other upstream materials. The underlying images remain governed by the original CUHK-PEDES access conditions and terms of use. Users must satisfy both the annotation license and the applicable upstream terms.

Citation

If you use MInterPEDES, please cite the ICML 2026 paper:

@inproceedings{bai2026interactive,
  title     = {Interactive Person Retrieval via Multi-Turn Multimodal Conversation},
  author    = {Bai, Yang and Wang, Tingfeng and Yang, Bin and Cao, Min and Wang, Jinqiao and Ye, Mang},
  booktitle = {Forty-third International Conference on Machine Learning},
  year      = {2026}
}

Because MInterPEDES builds on ChatPedes and CUHK-PEDES, please also cite the relevant upstream works:

@inproceedings{bai2025chat,
  author    = {Bai, Yang and Ji, Yucheng and Cao, Min and Wang, Jinqiao and Ye, Mang},
  title     = {Chat-based Person Retrieval via Dialogue-Refined Cross-Modal Alignment},
  booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages     = {3952--3962},
  year      = {2025}
}

@inproceedings{li2017person,
  author    = {Li, Shuang and Xiao, Tong and Li, Hongsheng and Zhou, Bolei and Yue, Dayu and Wang, Xiaogang},
  title     = {Person Search With Natural Language Description},
  booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages     = {1970--1979},
  year      = {2017}
}

Acknowledgements

MInterPEDES is built upon CUHK-PEDES. We thank the authors and maintainers of those resources and the related open-source projects.