FusionAgent: A Multimodal Agent with Dynamic Model Selection for Human Recognition
Paper • 2603.26908 • Published
You need to agree to share your Hugging Face information to access this dataset
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
This repository redistributes a PREPROCESSED version (HDF5 packed body crops and face crops) of the third-party MEVID person re-identification dataset, which is released by its original authors under a research-only license/agreement. By requesting access you confirm that (1) you have already obtained the original MEVID dataset from its official source and agreed to its terms, (2) you will use these files for non-commercial academic research only, (3) you will not redistribute them, and (4) you will cite the original dataset and our paper in any publication.
Log in or Sign Up to review the conditions and access this dataset content.
MEVID H5: preprocessed MEVID with body and face crops
Preprocessed HDF5 pack of the MEVID multi-view extended-video person re-identification benchmark, as used in our paper FusionAgent: A Multimodal Agent with Dynamic Model Selection for Human Recognition (arXiv:2603.26908, code).
| File | Size | Content |
|---|---|---|
mevid.h5.part00 … part16 |
361.7 GB | body frames, variable size (e.g. (100,40,3)); split into 20 GiB parts |
mevid_face_cleaned.h5.part00 … part02 |
55.4 GB | face crops, (112,112,3); split into 20 GiB parts |
train_name.txt, test_name.txt, track_train_info.txt, track_test_info.txt, query_IDX.txt |
~130 MB | official MARS-style split files, unchanged |
SHA256SUMS |
checksums of the two re-assembled H5 files |
mevid_face_cleaned.h5 is the face file after removing spurious/low-quality detections; it
is the version used in the paper.
Download and re-assemble
The two H5 files exceed the Hub's per-file limit and are split into 20 GiB chunks:
pip install -U huggingface_hub
hf auth login # you must have been granted access to this gated repo
hf download Paipile/mevid-h5 --repo-type dataset --local-dir ./data/mevid
cd ./data/mevid
cat mevid.h5.part* > mevid.h5 && rm mevid.h5.part*
cat mevid_face_cleaned.h5.part* > mevid_face_cleaned.h5 && rm mevid_face_cleaned.h5.part*
sha256sum -c SHA256SUMS
Point a MARS/MEVID-style loader at ./data/mevid. Split files: train_name.txt /
test_name.txt list one image name per line; track_*_info.txt rows are
start_idx end_idx pid outfit camid per tracklet; query_IDX.txt indexes into
track_test_info.txt.
H5 format
Keys mirror the directory structure of the original dataset; <img> is the original image
file name without extension. Every leaf is an HDF5 dataset holding one RGB uint8 array
of shape (H, W, 3) (i.e. np.asarray(PIL.Image.open(...).convert('RGB'))). Face crops are
aligned 112×112 crops detected inside each body image; face_<k> enumerates multiple
detections in the same image (most images have a single face_0). Images without a
detectable face are simply absent from the face file.
Open the files with h5py.File(path, 'r') and keep the handle open across __getitem__
calls (one handle per worker when using multiprocess DataLoader workers). The arrays are
stored uncompressed for fast random access.
| File | Key path | Shape |
|---|---|---|
mevid.h5 |
bbox_{train,test}/<pid>/<img> |
(H,W,3) |
mevid_face_cleaned.h5 |
bbox_{train,test}/<pid>/<img>/face_<k> |
(112,112,3) |
import h5py
from PIL import Image
body = h5py.File('data/mevid/mevid.h5', 'r'); face = h5py.File('data/mevid/mevid_face_cleaned.h5', 'r')
key = 'bbox_test/0201/0201O003C330T004F00006' # = bbox_test/0201/0201O003C330T004F00006.jpg
img = Image.fromarray(body[key][:])
if key in face:
faces = [Image.fromarray(face[key][k][:]) for k in face[key]]
License and terms
This repository contains derived data only; all rights to the images remain with the original authors. MEVID (Davila et al., MEVID: Multi-view Extended Videos with Identities for Video Person Re-Identification, WACV 2023) is derived from MEVA and released under a research-only license; see https://github.com/Kitware/MEVID. You must obtain the original dataset and accept its terms before using these files. Access here is gated: non-commercial academic research only, no redistribution.
Citation
If you use these files, please cite the original dataset above and our paper:
@article{zhu2026fusionagent,
title = {FusionAgent: A Multimodal Agent with Dynamic Model Selection for Human Recognition},
author = {Zhu, Jie and Guo, Xiao and Su, Yiyang and Jain, Anil and Liu, Xiaoming},
journal = {arXiv preprint arXiv:2603.26908},
year = {2026}
}
- Downloads last month
- 22