Examples, across cameras
Following people across eight cameras
Separate sightings become moves, moves become journeys, and the journeys sit on a site plan with the evidence for each step. This worked example was built by the CV-Scope developers in their own installation from the public MEVA dataset and is documented in the repository. It is not bundled with the code; the footage is available from mevadata.org.
-
G300 North walkway -
G419 Entrance stairs -
G420 Entrance doors -
G423 Upper hallway -
G330 Gym -
G638 Plaza and main entrance -
G424 Parking 2-3 -
G328 Parking 3
All eight cameras at the same moment, 11:22:40 on 11 March 2018. MEVA Known Facility KF1, Kitware and IARPA DIVA, CC BY 4.0.
The story
Eight synchronized cameras around a school building, 11 March 2018, 11:20-11:25. The gym is being cleared after a session (chairs are stacked at the stage end). Meanwhile the next groups arrive: two people walk along the north walkway and come in through the north doors, then a group of six and a group of four; they all go up the north staircase and along the upper hallway to the classrooms. Others cross the plaza and enter the main entrance. A maroon SUV drives around the parking lots.
Who was seen where, in order
From the documented cast. The numbers give the order of the sightings of each identity.
| Identity | G300 North walkway | G419 Entrance stairs | G420 Entrance doors | G423 Upper hallway | G424 Parking 2-3 | G328 Parking 3 | Notes |
|---|---|---|---|---|---|---|---|
| Visitor A | 1 | 2 | 3 | 4 | · | · | 3 moves; the first confirmed by the door sensor |
| Visitor B | · | · | · | · | · | · | walks with A |
| Visitor C | · | · | · | · | · | · | red top, in the group of six |
| Visitor D | · | · | · | · | · | · | straw hat, in the group of six |
| Visitor E | · | · | · | · | · | · | navy blazer, in the group of four |
| Visitor F | · | · | · | · | · | · | already upstairs: raises a restricted place without an entry alert |
| Maroon SUV | · | · | · | · | 1 | 2 | overlapping views |
What is real and what was added
The documentation separates measured, computed and annotated parts. Reproduced from it:
| Part | Source |
|---|---|
| Video | MEVA Known Facility (KF1), 5-minute clips of cameras G330, G423, G419, G420, G638, G424, G328, G300 (actors, public research dataset). Stored in data/videos/meva-kf1-2018-03-11/, remuxed from AVI to MP4 without re-encoding. |
| Detections, tracks, zone and line events, place relationships, "Walking together", "Parked" | Computed by CV-Scope: one run per camera (YOLO11x, ByteTrack, 10 fps). |
| Times | Each camera's Recorded at is the clip's start (11:20:00, 11:20:01, 11:20:04), dated today, so the cameras line up to the second. |
| Site plan | The aerial picture of the MEVA KF1 site map; 0.26 m per pixel. North is to the right. |
| Camera positions | Outdoor cameras from the MEVA camera models (KRTD: position, facing, field of view, GPS). Indoor cameras placed from the site map. |
| Travel times | Measured on the footage (north walkway → north doors about 20 s; north doors → upper hallway 1-9 s). |
| North door contact (sensor) | The door openings annotated in MEVA (person_opens_facility_door), replayed as sensor reports. |
| Identities (Visitor A-F, Maroon SUV) | Manual re-identification of the footage (clothing, build, timing), checked on crops of the frames. Stored as recognition events marked method: annotation; every link says "a reviewer identified", never "face recognition". The SUV's plate is a placeholder. |
A tour of the result in the application
- Locations › Topology. The hierarchy on the left (site › school › ground floor / upper floor › gym, entrance hall, stairs...). Click a camera to see its facing and field of view; click a link to see its travel time and the place on the way. Check a move: pick North walkway (G300) → Upper hallway (G423): "connected through 2 places", 8-100 s.
- Locations › Site view. Set the period to last 24 h. The alert on the right is Visitor F. Moves between cameras lists every move. Zoom into the school with the mouse wheel (labels keep their size). Type a visitor in Where was it last seen?.
- Relationships › Journeys, entity Visitor A. The timeline has 4 sightings and 3 moves; click the first move: Linked by: Manual annotation, 20 s against 8 s-1 min, Sensors: North door contact, confirmed 89 %. The location path shows the same steps on the plan.
- Relationships › Graph, around Visitor A. Try Replay in the time bar, 2 steps depth, and Path from here / to here between Visitor A and Visitor B.
- Relationships › Explorer and Timeline: the per-camera relationships behind all this (ENTERED, CROSSED, REMAINED_IN, TRAVELLED_WITH, PARKED_IN); open one and click Evidence.
- Analysis, run KF1 school 11:20 - Entrance doors (G420), Video review: the real video with the zones and the North doors doorway line; click an event to jump to it. The Crossing North doors reverse events are people coming in (doorway line, "appeared").
- Recognition › Recognition events: the 19 annotation events, with a crop of each sighting.
How moves are scored
Each move between cameras is compared with the travel time the location model expects, checked against the camera topology, weighted by how the identity was established and by sensors on the way, and given a state: confirmed, likely or possible. Nothing turns a likely move into a fact.
MEVA (Multiview Extended Video with Activities), Known Facility KF1, Kitware and IARPA DIVA, released under CC BY 4.0: <https://mevadata.org>. K. Corona et al., "MEVA: A Large-Scale Multiview, Multimodal Video Dataset for Activity Detection", WACV 2021. Site map v1.3 (2019) and camera models: <https://gitlab.kitware.com/meva/meva-data-repo>.