Sample data

Three files, all served from this domain, no auth:

$ curl -O https://prxintel.com/challenge/clip.mp4        # the footage, 15.6 s, 720p
$ curl -O https://prxintel.com/challenge/zone.json       # zone geometry + entry events
$ curl -O https://prxintel.com/challenge/detections.json # every tracked box, every frame

zone.json#

{
  "fps": 29.97,
  "clipSeconds": 15.58,
  "zoneFrames": [ [[x, y], [x, y], [x, y], [x, y]], ... ],
  "events":     [ { "t": 0.067, "x": 0.54, "y": 0.909, "id": 7 }, ... ]
}

zoneFrames holds one quadrilateral per frame, 467 in total. This is the world-anchored zone polygon as registered on each individual frame; play the clip and draw frame i's quad over it and you'll see the zone hold its patch of road while the camera moves. events holds the 15 scored entries: time in seconds, normalized entry position, and the track ID of the car.

detections.json#

{
  "fps": 29.97,
  "frames": [ [[x, y, w, h, track_id], ...], ... ]
}

One array per frame, one 5-tuple per tracked object: normalized box plus its persistent track ID. This is what the challenge replay draws. Short-lived and implausibly small tracks are filtered out (minimum 8 frames of life, minimum median area), everything else is left as the tracker produced it, including its mistakes.

Check us#

The claim the challenge makes is that 15 distinct cars enter the zone. You don't have to take that on faith:

import json, urllib.request

zone = json.load(urllib.request.urlopen("https://prxintel.com/challenge/zone.json"))
print(len(zone["events"]), "entries")
for ev in zone["events"]:
    print(f'  {ev["t"]:6.2f}s  track {ev["id"]:3d}  at ({ev["x"]:.2f}, {ev["y"]:.2f})')