Voxel51/extended_video_activities_drop_4

Dataset Card for meva_mevid

[image]

This is a FiftyOne dataset with 201 samples.

Installation

If you haven't already, install FiftyOne:

pip install -U fiftyone

Usage

import fiftyone as fo
from fiftyone.utils.huggingface import load_from_hub

# Load the dataset
# Note: other available arguments include 'max_samples', etc
dataset = load_from_hub("Voxel51/extended_video_activities_drop_4")

# Launch the App
session = fo.launch_app(dataset)

Dataset Card — MEVA Drop-4 / MeVID (meva_mevid)

MEVA drop-4-hadcv22 is a 201-clip subset of the Multiview Extended Video with Activities (MEVA) Known Facility 1 dataset. It was released specifically to support the HADCV22 Self-Reported Leaderboard Challenge and the broader ActEV activity-detection evaluation series run by NIST. The MeVID (Multi-view Extended Video ID) person re-identification benchmark builds its test/query annotations on top of this same footage.

Dataset Details

Collection

The footage was recorded at the Muscatatuck Urban Training Center (MUTC) with over 100 actors performing scripted scenarios. It was collected by Kitware Inc. under the IARPA Deep Intermodal Video Analytics (DIVA) program.

Annotations (MeVID)

Person re-identification tracklets from the MeVID benchmark. Each tracklet records that a specific person (identified by a cross-camera integer ID) was visible in a specific clip for a given number of consecutive frames.

317 of the 499 test tracklets are designated query tracklets for the re-identification evaluation (matching a query sequence against a gallery of test sequences across cameras and time).

Coverage: drop-4 accounts for roughly 14% of the full MeVID annotation. The remaining 86% of MeVID tracklets reference cameras from other MEVA drops not included in this download and are silently ignored by the parser.

The MeVID annotation provides tracklet-level metadata only — it does not include per-frame bounding boxes, absolute frame offsets within a clip, or activity labels. The MEVA dataset does publish separate DIVA-format activity annotations with bounding boxes (available via the MEVA data repo), but those are a separate download not included here.

FiftyOne Dataset Structure

Each of the 201 samples corresponds to one 5-minute MP4 clip.

Sample-level fields

Tracklet dict schema

Each entry in the tracklets list describes one contiguous person track:

Filtering examples

import fiftyone as fo

ds = fo.load_dataset("meva_mevid")

# All test clips that contain a query tracklet
queries = ds.match(
    (fo.ViewField("split") == "test") & fo.ViewField("has_query")
)

# Hospital clips only
hospital = ds.match(fo.ViewField("scene") == "hospital")

# Clips from a specific camera
cam424 = ds.match(fo.ViewField("camera_name") == "G424")

# Clips where person 212 appears
person212 = ds.match(fo.ViewField("person_ids").contains(212))

Source & License

• —Videos: mevadata.org — licensed under CC BY 4.0
• —MeVID annotations: distributed with the MeVID benchmark
• —Produced by: Kitware Inc. and IARPA

Citation

@InProceedings{Corona_2021_WACV,
    author    = {Corona, Kellie and Osterdahl, Katie and Collins, Roderic and Hoogs, Anthony},
    title     = {MEVA: A Large-Scale Multiview, Multimodal Video Dataset for Activity Detection},
    booktitle = {Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV)},
    month     = {January},
    year      = {2021},
    pages     = {1060-1068}
}