younght/MPEDES

M-PEDES

M-PEDES is an evaluation-only benchmark for multilingual text-based person retrieval (M-TBPR). It extends the test sets of CUHK-PEDES, ICFG-PEDES, and RSTPReid from English to Arabic, German, Spanish, French, Japanese, Korean, Russian, and Chinese.

The repository contains translated annotations only. It does not include training data, original English captions, or original pedestrian images.

Dataset Summary

Source dataset	Queries per language	Target languages
CUHK-PEDES	6,156	8
ICFG-PEDES	19,848	8
RSTPReid	2,000	8
Total	28,004	8

In total, M-PEDES contains 224,032 translated queries. The translations were generated with GPT-5.2 using retrieval-aware prompts that emphasize the preservation of fine-grained visual attributes. A 10% stratified sample was manually audited. Across the full benchmark, the average cross-lingual semantic similarity is 0.891, back-translation BLEU is 29.6, and attribute preservation is 97.6%. Full construction, quality-control, and evaluation details are provided in the accompanying paper.

Data Format

Each configuration follows {source-dataset}-{language} and contains one test split. For example, cuhk-pedes-zh provides:

{
  "id": 12004,
  "file_path": "train_query/p8848_s17661.jpg",
  "caption_idx": 0,
  "caption_zh": "一个穿着蓝白条纹背心、绿色裤子和粉色鞋子的男人。"
}

caption_XX denotes the translated description, where XX is the target-language code.

Usage

from datasets import load_dataset

dataset = load_dataset(
    "younght/MPEDES",
    "cuhk-pedes-zh",
    split="test",
)
print(dataset[0])

Users must obtain the corresponding source images from the original dataset providers and resolve each record's file_path relative to the appropriate dataset root.

Intended Use

M-PEDES is intended for evaluating multilingual and cross-lingual text-based person retrieval. Results should be reported separately for each source dataset and language using the evaluation protocol described in the accompanying paper.

Limitations and License

The captions are machine-generated translations and may retain noise or ambiguity from the original English annotations.

The multilingual translations and original documentation are released under the M-PEDES Research-Only License v1.0 for non-commercial academic research and educational use. Source images, English captions, identifiers, paths, and other upstream materials are not licensed by M-PEDES and remain subject to their original terms. Users must obtain the source datasets independently from their official providers.

Citation

If you use M-PEDES in your research, please cite:

@inproceedings{yang2026gram,
  author    = {Yang, Haotian and Peng, Cheng and Wang, Haobo and Zhu, Zhen and Tang, Xiu and Yuan, Gongsheng and Xie, Zhongle and Wu, Sai and Wang, Weiqiang and Cheng, Yu},
  title     = {Beyond Contrastive: Generative Reranking for Multilingual Text-Based Person Retrieval},
  booktitle = {Proceedings of the 34th ACM International Conference on Multimedia},
  series    = {MM '26},
  year      = {2026},
  month     = nov,
  location  = {Rio de Janeiro, Brazil},
  doi       = {10.1145/3767308.3835886},
  isbn      = {979-8-4007-2213-4/2026/11},
  publisher = {Association for Computing Machinery},
  license   = {CC BY 4.0}
}

Please also cite the original CUHK-PEDES, ICFG-PEDES, and RSTPReid papers.