Travor278/whoego-topview

Who-is-Ego top-view queries

Each query pairs a top view (drone or overhead camera, the exo view) with an ego view taken at the same instant by a camera worn by one of the people (or mounted on one of the vehicles) in the top view. The task is to say which candidate carries the ego camera. Gallery: Travor278/whoego-topview-gallery.

Only the frames the queries use are included, byte for byte as in the source: no re-encoding, plain uncompressed tars. The source annotation files come along unchanged. Every subset keeps the terms of its source (see Sources and terms).

subset	manifest	queries	candidates per query, mean (min-max)	random	wearer label
csrd_virtual	csrd_virtual.jsonl	4,008	18.9 (10-25)	0.056	exact (synthetic)
csrd_virtual	csrd_virtual_ref.jsonl	1,002	18.9 (10-25)	0.056	exact, but see below
csrd_real	csrd_real_s5.jsonl	2,309	9.4 (8-12)	0.107	inferred, see below
cvmhat	cvmhat_s10.jsonl	1,170	9.0	0.11	top-view ID 0 (dataset convention)
airv2x_mini	airv2x_drone.jsonl / airv2x_drone_boxes.jsonl	214	23.7	0.046	exact (synthetic, vehicles)
airv2x_mini	airv2x_rsu.jsonl	10			exact (synthetic, vehicles)

Layout and format

<subset>/<manifest>.jsonl        one query per line
<subset>/frames_*.tar            images, unchanged; extract inside <subset>/ (tar xf frames_X.tar -C <subset>)
<subset>/source_annotation/      the source's annotation files, unchanged
csrd_real/wearers.json           evidence for every inferred wearer

A query line:

• exo, ego: image paths relative to the subset directory.
• cand: candidate points [u, v] in exo pixels: the centre of each person's box in the top view, or the mounting point of the ego camera on each vehicle.
• label: the index of the carrier in cand.
• key: <set>/<group>/<frame>/<view>. All queries with the same top frame share the prefix, so their wearers can be assigned jointly.
• boxes: optional, [x0, y0, x1, y1] per candidate.
• ids: optional, the source dataset's IDs.

Candidates are shuffled with a fixed seed per query, so the position of the answer carries no information. AirV2X queries also give K, w2c_exo and w2c_ego (OpenCV world-to-camera). The other subsets are uncalibrated.

Notes per subset

• csrd_virtual (CSRD-II / CSRD-V): 1002 independent synthetic frames, each with one top view at 21.6 m and five horizontal views at 1.8 m (1024 x 768 PNGs). source_annotation/annotation/ holds the source's camera and person positions, top-view boxes with pids and per-view boxes (txt / pth).

  • The camera of view k is carried by pid 20 + k: in every frame, camera<k>.txt places the camera at that pid's position (distance 0).
  • Camera 1 is the reference frame of the source dataset. It sits at the origin, facing 0 deg, in every frame, so its wearer is on the same top-view pixel in all 1002 frames. A position prior answers those queries, so they are kept apart in csrd_virtual_ref.jsonl.

• csrd_real (CSRD-R): 7 groups, each a drone top view plus 2-4 head-mounted GoPros, with boxes and person IDs that are shared between the views. The source does not label who wears which camera. We infer it per camera from two cues:

  • the wearer is never boxed in their own view;
  • from the wearer's top-view position, the bearings of the people boxed in the ego view run in the same left-to-right order as their boxes (Kendall tau at the best heading).

  15 of 19 cameras pass all of these: tau >= 0.95, never in their own view, and a lead of >= 0.2 over the next candidate. The kept cameras score tau 0.97-1.00. Per query, the median tau is 1.00 and only 4 of 2,308 queries are below 0.5. The wearers' tracks never jump, so there are no ID switches. 2688 x 1512 JPEGs as published, every 5th frame; source_annotation/GT_txt/ holds the source boxes of every view.

• cvmhat: 15 groups, one drone top view plus one head-mounted GoPro each. The wearer is ID 0 in the top-view ground truth (dataset convention). Every 10th frame, 2688 x 1512 JPEGs as published; source_annotation/ holds the source's GT_txt.zip and Tra_Gopro.zip. CSRD-R comes from the same group's recordings, so treat csrd_real and cvmhat as one real-data source, not two.

• airv2x_mini: CARLA vehicles from the AirV2X-Perception mini split.

  • ego: a vehicle's front camera;
  • exo: a drone's nadir camera (drone) or a road-side unit camera (rsu);
  • candidates: every visible vehicle, with the ego camera's mounting point placed on each;
  • airv2x_drone_boxes.jsonl is the same set of queries with candidate boxes added.
  • The frames tar keeps the source layout (test/timestamp_*/agent_*/) and includes each used agent's metadata.pkl and each timestamp's objects.pkl.

Sources and terms

• CSRD: Qian, Han, Feng, Wang, From a Bird's Eye View to See: Joint Camera and Subject Registration without the Camera Calibration, CVPR 2024 (BEVSee, HF clarkqian/CSRD). The source states no licence. Wearer labels for CSRD-R are ours.
• CVMHAT: Han et al., Complementary-View Multiple Human Tracking, AAAI 2020, and Multiple Human Association and Tracking from Egocentric and Complementary Top Views, TPAMI (RuizeHan/CVMHAT). For academic research only, per the source.
• AirV2X-Perception: xiangbog/AirV2X-Perception, CC BY 4.0.

Please cite the original works when you use a subset. If you are a rights holder and want a subset removed, open a discussion on this repository.