HugC/MVEB-train

MVEB Train Split

English | 简体中文

This repository contains the train split of MVEB (Multimodal Visual identity Embedding Benchmark) — a benchmark for identity-level retrieval. Given a query (text + image), a model retrieves candidates (text + image) that belong to the same identity.

The train split covers 20 subsets across four meta-tasks:

• Identity Recognition — object / product / species recognition
• Re-Identification — person / face / vehicle re-ID
• Identity Grounding — grounding queries to visual identities
• Identity Editing — retrieval over edited image pairs

The test split is published separately and includes 8 additional OOD subsets not present in train (e.g. Product1m, Market1501, FORB, OpenGPT4o). See the corresponding test repository for evaluation data.

Not every subset is redistributable. CompCars, DukeMTMC, Inshop, IUST, MS-Celeb-1M and MultiID are not shipped in this repository — rebuild them locally as described in Rebuilding Restricted Subsets. The other 14 subsets are included as parquet files.

Directory Layout

Each subset is stored as three components:

File	Description
query.parquet	Query metadata (instruction, text, pos_ids, …)
candidate.parquet	Candidate-pool metadata
media-*.parquet	Deduplicated media pool (image / video / audio)

query and candidate rows reference the shared media pool via media_index, so each image is stored only once.

Loading Data

Each subset exposes three HuggingFace configs: {Subset}_media, {Subset}_query, and {Subset}_candidate. The split name is train.

from datasets import load_dataset

repo = "HugC/MVEB-train"  # Hub repo id, or a local directory path

query = load_dataset(repo, "Cars196_query", split="train")
candidate = load_dataset(repo, "Cars196_candidate", split="train")
media = load_dataset(repo, "Cars196_media", split="train")

print(query[0])
# {'id': '0', 'instance_id': '0', 'instruction': '...', 'text': '...',
#  'pos_ids': ['1', '2', ...], 'media_index': 0}

Note: The third positional argument to load_dataset is data_dir, not split. Always pass split="train" as a keyword argument.

Resolving Images

query and candidate parquets do not embed images. Resolve them through media_index:

from datasets import load_dataset

repo = "HugC/MVEB-train"
query = load_dataset(repo, "Cars196_query", split="train")
media = load_dataset(repo, "Cars196_media", split="train")

row = query[0]
image = media[row["media_index"]]["image"]  # PIL.Image

Schema

query.parquet (typical fields):

Field	Type	Description
id	string	Query sample id
instance_id	string	Identity / instance id (absent in some subsets)
instruction	string	System instruction
text	string	User text
pos_ids	list[string]	Positive candidate id list
media_index	int	Row index into the media pool

candidate.parquet — similar to query; usually without pos_ids.

media-*.parquet — columns image, video, audio. Currently image-only; video and audio are reserved for future multimodal extensions.

Rebuilding Restricted Subsets

A few subsets come from sources whose licences do not allow redistribution, so they are not shipped in either repository and must be rebuilt locally from the original providers. scripts/ holds that pipeline for both splits — it is shipped only in this repository, and the same scripts also produce the corresponding test subsets.

Covered subsets: CompCars, DukeMTMC, Inshop, IUST, MS-Celeb-1M, MultiID (train + test) and Product1m (test only).

Layout

scripts/
├── run.sh                  # one-click runner for all subsets
├── pack_media_parquet.py   # shared parquet packing library
└── <Subset>/
    ├── process.sh           # download -> extract -> pack
    ├── process_<subset>.py   # split-aware parquet builder
    └── train_test_split.json # frozen split, keeps rebuilds reproducible

Recommended Order

Download both repositories before running the scripts, so that rebuilt subsets land next to the already published ones:

huggingface-cli download HugC/MVEB-train --repo-type dataset --local-dir MVEB-train
huggingface-cli download HugC/MVEB-test  --repo-type dataset --local-dir MVEB-test

bash MVEB-train/scripts/run.sh              # all subsets
bash MVEB-train/scripts/run.sh DukeMTMC     # selected subsets
bash MVEB-train/scripts/run.sh --list       # show available subsets

Running the scripts first and downloading afterwards is discouraged: the download never removes local files, so stale media-*.parquet shards from an earlier rebuild could be picked up alongside the downloaded ones.

Paths

With the layout above no configuration is needed. Override these environment variables to place outputs elsewhere:

Variable	Meaning	Default
MVEB_TRAIN_DIR	destination for train subsets	this repository (parent of scripts/)
MVEB_TEST_DIR	destination for test subsets	MVEB-test next to this repository
MVEB_ROOT	scratch space for downloads/, source/, logs/	parent of this repository

MVEB_TEST_DIR=/data/MVEB-test MVEB_ROOT=/scratch bash MVEB-train/scripts/run.sh

Raw downloads and extracted images under MVEB_ROOT are only needed while rebuilding and can be deleted afterwards.

Requirements

• python3 with datasets, pyarrow, pillow, tqdm, requests
• kaggle (DukeMTMC, Inshop, IUST) with an API token in ~/.kaggle/kaggle.json
• huggingface-cli (MS-Celeb-1M, MultiID), gdown (CompCars), git (Product1m)
• unzip, zip, tar

run.sh reports missing commands before it starts.

Re-running After a Failure

Every stage is resumable, so simply re-run the same command:

• downloads resume or skip already fetched files; Product1m re-fetches only missing images
• extraction is guarded by an .extracted marker written on success, so an interrupted archive is redone
• parquet packing always rewrites the subset directory from scratch

A failing subset does not abort the others. Per-subset logs are written to ${MVEB_ROOT}/logs/<Subset>.log, and the final summary lists the command that retries just the failed subsets.

If CompCars reports a truncated split volume, delete the file it names and re-run — gdown skips existing files by size-agnostic name matching, so a partial volume has to be removed explicitly.

Citation

If you find MVEB useful in your research, please cite:

@inproceedings{cao2026illuminating,
  title={Illuminating Visual Identity in Universal Multimodal Embeddings},
  author={Cao, Jiawei and Feng, Junyi and Hua, Jiashen and Huang, Ziheng and Deng, Bing and Wu, Kaijie and Gu, Chaochen and Ye, Jieping},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={8737--8748},
  year={2026}
}