Voxel51/extended_video_activities_drop_4

Dataset Card for meva_mevid

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.

Property	Value
Clips	201
Total duration	~16.75 hours
Clip length	~5 minutes each
Resolution	1920 × 1080
Frame rate	30 fps
Codec	H.264 (remuxed to MP4 for FiftyOne)
Cameras	19 fixed ground cameras
Recording dates	7 days (March–May 2018)
Scenes	hospital, school, bus

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.

Split	Clips with tracklets	Tracklets	Unique persons
train	77	657	—
test	64	499	—
unlabeled	60	0	—
total	141	1 156	155

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

Field	Type	Description
filepath	str	Absolute path to the .mp4 file
recording_date	str	Recording date, e.g. "2018-03-12"
start_time	str	Clip start time, e.g. "12:20:04"
end_time	str	Clip end time, e.g. "12:25:04"
scene	str	Camera location: "school", "hospital", or "bus"
camera_id	int	Numeric camera ID, e.g. 424
camera_name	str	Camera label used in MEVA filenames, e.g. "G424"
temporal_segment	int	0-based index of this clip within its camera's chronological sequence (matches the T value in MeVID filenames)
split	str	"train", "test", or "unlabeled"
has_query	bool	True if at least one query tracklet originates from this clip
person_ids	list[int]	Sorted list of unique person IDs visible in this clip per MeVID
num_tracklets	int	Total annotated tracklets in this clip
tracklets	list[dict]	One dict per tracklet — see below

Tracklet dict schema

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

Key	Type	Description
person_id	int	Cross-camera person identity (shared across all clips)
occurrence_id	int	Which occurrence this is for this person (a person may appear multiple times across the dataset)
frame_count	int	Number of frames in the track
split	str	"train" or "test"
is_query	bool	Whether this tracklet is a re-ID query

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}
}