TVPReid (unofficial Hub mirror)

This repository is a convenience mirror for research. It is not an official release by the TVPReid authors.

The uploader (bassatbassat) does not claim copyright over the videos, captions, or dataset design. All credit belongs to the original authors listed below. If you use this data, cite their papers, follow their licenses, and prefer the official distribution when possible.

If you are an original rights holder and want this mirror taken down or gated further, open a Hub discussion or contact the uploader.

---

What this dataset is

TVPReid is a text-to-video person retrieval benchmark: given a natural-language description, retrieve the matching pedestrian video.

It was introduced by Zhang et al. at ACM Multimedia 2024. Videos were assembled from three existing person re-identification datasets and each video was annotated with two English captions (appearance, motion, and scene).

Average caption length is about 30 words (longest 83), according to the official README. The Hub copy matches the updated TVPReid release (the authors note it differs from some numbers in the original paper).

Official project: NjtechCVLab/TVPReid-Dataset
Official paper: TVPR: Text-to-Video Person Retrieval and a New Benchmark (arXiv:2307.07184)

---

Required citations

1. TVPReid / TVPR (captions, splits, and this benchmark)

@inproceedings{zhang2024tvpr,
  title     = {TVPR: Text-to-Video Person Retrieval and a New Benchmark},
  author    = {Zhang, Xu and Ni, Fan and Dong, Guan-Nan and Zhu, Aichun
               and Wu, Jianhui and Ni, Mingcheng and Liu, Hui},
  booktitle = {Proceedings of the 32nd ACM International Conference on Multimedia},
  pages     = {10105--10113},
  year      = {2024}
}

2. Source video datasets (you must also cite the subset you use)

PRID-2011 (prid) — Hirzer et al., SCIA 2011:

@inproceedings{hirzer2011prid,
  title     = {Person Re-Identification by Descriptive and Discriminative Classification},
  author    = {Hirzer, Martin and Beleznai, Csaba and Roth, Peter M. and Bischof, Horst},
  booktitle = {Scandinavian Conference on Image Analysis (SCIA)},
  pages     = {91--102},
  year      = {2011}
}

iLIDS-VID (ilids) — Wang et al., ECCV 2014:

@inproceedings{wang2014ilidsvid,
  title     = {Person Re-Identification by Video Ranking},
  author    = {Wang, Taiqing and Gong, Shaogang and Zhu, Xiatian and Wang, Shengjin},
  booktitle = {European Conference on Computer Vision (ECCV)},
  pages     = {688--703},
  year      = {2014}
}

DukeMTMC-VideoReID (duke) — Wu et al., CVPR 2018, built on DukeMTMC (Ristani et al., ECCV 2016):

@inproceedings{wu2018dukevideoreid,
  title     = {Exploit the Unknown Gradually: One-Shot Video-Based Person Re-Identification by Stepwise Learning},
  author    = {Wu, Yu and Lin, Yannis and Dong, Xu and Yan, Yan and Ouyang, Wanli and Yang, Yi},
  booktitle = {Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
  pages     = {5177--5186},
  year      = {2018}
}

@inproceedings{ristani2016dukemtmc,
  title     = {{Performance Measures and a Data Set for Multi-Target, Multi-Camera Tracking}},
  author    = {Ristani, Ergys and Solera, Francesco and Zou, Roger and Cucchiara, Rita and Tomasi, Carlo},
  booktitle = {European Conference on Computer Vision Workshop (ECCVW)},
  year      = {2016}
}

---

Copyright, license, and ethics

• License on this Hub repo: other. There is no license granted by the uploader. Redistribution here does not replace the original terms.
• TVPReid’s official GitHub states that users download third-party datasets and must check those projects’ license terms before use.
• PRID-2011 and iLIDS-VID are typically released for academic research by their original institutions. Confirm terms with the original distributors before any use beyond research.
• DukeMTMC was withdrawn by Duke University after privacy and consent concerns about campus surveillance footage. duke in this mirror is derived from DukeMTMC-VideoReID. Prefer prid and ilids if you can avoid Duke-derived data. Do not use any split to identify real people.
• Videos show real pedestrians in surveillance settings. This is sensitive personal data. Do not use it for commercial products, law-enforcement deployment, or deanonymization.
• This mirror exists so SHAWAF / other research code can load_dataset without Google Drive. It is not an assertion of ownership.

---

Dataset structure

TVPReid-PRID/   captions/ + videos/personXXXX.mp4
TVPReid-iLIDs/  captions/ + videos/personXXXX.mp4
TVPReid-Duke/   captions/ + videos/personXXXX.mp4
metadata/       Hub split tables (JSONL)

Each metadata row:

---

Load on the Hub

from datasets import load_dataset

# Smallest subset (good for a pipeline smoke test)
prid = load_dataset("bassatbassat/TVPReid", "prid")

# Full benchmark
all_splits = load_dataset("bassatbassat/TVPReid", "all")

row = prid["test"][0]
print(row["video_id"], row["captions"][0])
# Decode the video file from the Hub snapshot:
# datasets.Video() can be applied after download if you need frames.

After load_dataset(..., download=True) the video column is a relative path inside the snapshot. Point your dataloader at:

from pathlib import Path
from datasets import load_dataset, DownloadMode

ds = load_dataset("bassatbassat/TVPReid", "prid", split="test")
# Hub copies files next to the JSONL in the dataset cache / repo snapshot.

Official captions and train/val/test IDs are also kept under each subset’s captions/ folder (TVPReid-*.json, train.csv, val.csv, test.csv).

---

Intended use

• Research on text-to-video person retrieval / text-to-tracklet search
• Frozen VLM baselines (X-CLIP, LanguageBind, InternVideo2-CLIP, etc.)

Not intended for: identifying individuals, operational surveillance, or commercial biometric systems.

---

Acknowledgement

Videos and labels come from:

1. Zhang, Ni, Dong, Zhu, Wu, Ni, and Liu — TVPReid / TVPR (ACM MM 2024)
2. Hirzer, Beleznai, Roth, and Bischof — PRID-2011 (SCIA 2011)
3. Wang, Gong, Zhu, and Wang — iLIDS-VID (ECCV 2014)
4. Wu, Lin, Dong, Yan, Ouyang, and Yang — DukeMTMC-VideoReID (CVPR 2018)
5. Ristani, Solera, Zou, Cucchiara, and Tomasi — DukeMTMC (ECCV 2016)

This Hub snapshot was prepared for the SHAWAF graduation project as an unofficial research mirror of the files published at NjtechCVLab/TVPReid-Dataset.