Travor278/whoego-synth

Who-is-Ego synthetic ego/exo episodes

This is synthetic multi-person indoor data for ego/exo wearer identification: given an egocentric frame, find which person in an exocentric view is wearing that camera. Each episode is one static moment in a procedurally generated room. Every person wears the same camera props (glasses on the head, a chest camera), so the chest prop itself gives no shortcut. Every label comes directly from how the scene was built.

Review galleries (private, ask for access):

• people v1: https://huggingface.co/spaces/Travor278/whoego-people-v1-gallery
• outdoor people v1 / v2: https://huggingface.co/spaces/Travor278/whoego-outdoor-gallery
• people v2: https://huggingface.co/spaces/Travor278/whoego-people-v2-gallery
• v0 sets: https://huggingface.co/spaces/Travor278/whoego-synth-gallery

Versions (2026-11-02)

• v1 (people_v1/): people up to the realism standard (BEDLAM bodies and real motion-capture poses, mixed facing), rectilinear head and chest cameras.
• v2 (people_v2/, 2026-11-02): the v1 pipeline with mixed real-device lenses (action-camera fisheye / wide, camera glasses, body cams), children and teenagers, larger crowds and everyday activities; 8 episodes per room.
• v0: the earlier MPFB sets, kept for reference: v0/ (the former data/ and v1/ folders) and outdoor_v0/ (the former outdoor/). v0/v1_<n>.tar is still the fixed synthetic test set used in our earlier results.

people v1 (people_v1/, uploading since 2026-11-02)

• Rooms: the v0 room generators (Infinigen home rooms, 28–60 m² living and dining rooms, meeting rooms, offices and labs); 269 rooms x 2 episodes = 537 episodes, up to 8 people each (2,636 people).
• People: 70 % SMPL-X bodies with BEDLAM skin and clothing textures (only textures that cover the belly; no swimwear, no tops close to skin colour), posed with real BEDLAM motion-capture frames (7,475 standing, 226 sitting); 30 % MPFB bodies. Hair and shoes come from same-sex MPFB characters.
• Facing: talking to the nearest person 20 %, towards the group 25 %, looking at something in the room 1.2–5 m away 30 %, any direction with at least 1 m free in front 25 % (often back or side to the others); seated people face their seat. Every person has a facing field.
• Views: 1–2 exo (1280x720, tripod or wall cameras), per person a head camera (704x704, about 100 deg) and a chest camera (960x540, about 110 deg), all rectilinear. 1,034 exo, 2,530 head and 2,164 chest views after QA.
• Imaging: <view>.png passed through a camera ISP model (exo: tripod / wall camera; ego: small-sensor action camera; noise, white balance, tone curve, JPEG; parameters in isp.json); <view>_clean.png is the render before the ISP.
• QA: ego views facing a near surface (median depth < 0.7 m, e.g. a seated chest camera that only sees the legs and the table), exo views blocked by a near object, and ego views mostly covered by the wearer's own body are dropped (dropped_views).
• Licence note: SMPL-X, BEDLAM and the Meshcapade textures are for non-commercial research only.

people v2 (people_v2/, 2026-11-02)

The people v1 pipeline (the same 269 rooms, BEDLAM + MPFB people, mixed facing, camera ISP, 32 samples), made larger and closer to real recordings: real-device lenses, children and teenagers, gatherings and everyday activities.

• Episodes: 8 per room; every episode draws new people, positions, cameras and lenses. 2,133 episodes in 269 rooms: 80 std home rooms (634 episodes), 30 big home rooms (238) and 159 work rooms (1,261). A few draws whose exo cameras found no position were skipped.
• Lenses (datagen/synth/lenses.py): each view is rendered as a wide equidistant image and remapped exactly to the device model.
  • head: Aria-like circular fisheye 45 % (aria, 1024², 105–140°), GoPro on a head strap 20 % (gopro_head), camera glasses 15 % (smartglasses, 720×960 portrait), headset camera 10 % (headset), wide rectilinear 10 % (pinhole_wide, 704²);
  • chest: GoPro wide 30 % (gopro_wide), the TF2023 chest camera model 25 % (tf_chest, 960×540), body cam 15 % (bodycam), rectilinear 30 % (linear, 960×540);
  • exo (1280×720, tripod or wall): rectilinear 70 %, GoPro fisheye 20 % (gopro_exo), wide CCTV 10 % (cctv_wide);
  • in about half of the episodes all wearers share one head and one chest device (lens_scheme), as a team issued the same hardware would.
  • Each view records camera_model (PINHOLE / OPENCV / OPENCV_FISHEYE), K, dist, device, hfov and, for circular fisheyes, valid_radius.
• People (11,840, 5.55 per episode, at most 39; 7,840 BEDLAM, 4,000 MPFB):
  • children and teenagers, 6–17 years (MPFB bodies; about 12 % of the people in home rooms, fewer in work rooms; child, age_years): 637;
  • BEDLAM people who stand take an everyday activity from the BEDLAM motion-capture frames: standing 45 %, walking 20 %, gesturing 10 %, a phone call 10 %, holding a cup, a phone or a folder 15 % (a prop in the hand; activity, prop). Counts: standing 5,102 (BEDLAM and MPFB), sitting 3,824, walking 1,070, holding something 816, gesturing 532, phone call 496;
  • gatherings (crowd): where the free floor allows it, 35 % of the episodes are a party, a meeting or a class break of 12–25 people (about one person per 1.5 m²): 172 episodes. The other episodes have 2–8 people, except 29 (1.4 %, crowd false) with 9–39; use the length of people for the group size.
• Camera heights (world z of the ego cameras): 0.57–1.86 m, 5–95 % 0.94–1.67 m, median 1.34 m; 25 % below 1.2 m (seated people and children).
• Views: 4,075 exo, 11,390 head and 9,609 chest views after QA. Ego views per device: aria 5,269, gopro_wide 2,952, linear 2,796, tf_chest 2,433, gopro_head 2,073, smartglasses 1,738, bodycam 1,428, pinhole_wide 1,196, headset 1,114.
• Queries (as in our loader: an exo view with at least 2 people of ≥ 400 px, times each ego view whose wearer is one of them; candidates = those people): 37,783 (20,559 head, 17,224 chest), 7.09 candidates on average, random baseline 0.192.
• QA: at render time, ego views facing a near surface (walls and furniture only: people close to the camera in a crowd are normal), exo views blocked by a near object and ego views mostly covered by the wearer's own body are dropped; after rendering, view_qa.py (dark, lens inside a surface, own body > 50 %, near surface, blocked exo, one person over 35 % of an exo view). All drops are in dropped_views: ego 2,681 of 23,680 (11.3 %: near surface 2,569, own body 111, dark 1), exo 170 of 4,245 (4.0 %, blocked). Contact sheets of exo and ego views were reviewed by eye before publication.
• Known limitation: the BEDLAM pose library is the one people v1 uses. About 18 % of its seated poses have both hands above the shoulders (cheering, stretching) and about 6 % of its plain standing poses have both arms spread, so a few percent of the people strike such a pose.
• Layout: as people v1: people_v2/{std,big,work}_<seed>.tar holds <room>/summary.json and <room>/ep00<k>/. Views add the lens fields above; people add child, age_years, activity, prop; episodes add crowd and lens_scheme. Counts and breakdowns: people_v2/stats.json.
• Camera points: people_v2/cam_points.jsonl (schema whoego.cam_points/2, see "Camera points" below).
• Generator: datagen/synth/compose.py --body mix --p-bedlam 0.7 --isp --max-people 8 --lens mix --activities --p-child 0.12 --p-crowd 0.35 --crowd-max 25 (batch synth_batch.sh, 32 samples).

outdoor people v1 / v2 (outdoor_v1/, outdoor_v2/, 2026-11-03)

People outdoors on Infinigen nature terrain (high-quality terrain: forests, a grassy plain, rivers, mountains, a canyon, a coast). It replaces the flat ground and HDRI sky of outdoor_v0/. Each episode is one still moment, built with the people v1 pipeline.

• Versions: same scenes and generator, independent episodes.
  • outdoor_v1/: rectilinear head and chest cameras.
  • outdoor_v2/: mixed real-device lenses (datagen/synth/lenses.py): Aria-like circular fisheye, GoPro, camera glasses and body cams on the people; the exo cameras are partly GoPro fisheye or wide CCTV. Each view records its camera_model (PINHOLE / OPENCV / OPENCV_FISHEYE), dist, device and, for circular fisheyes, valid_radius. lens_scheme says whether all wearers of the episode share one device.
• Scenes: the spots stand 4–45 m in front of Infinigen's own camera, the only part of a nature scene that is finished (fine terrain, trees, bushes). Arctic, snow and desert scenes are left out (city clothes look wrong there).
• People:
  • adults: 70 % SMPL-X bodies with BEDLAM textures, 30 % MPFB bodies;
  • children (7–11 years) and teenagers (12–16): MPFB, in the crowd scenarios;
  • BEDLAM adults take an everyday pose of their activity: hikers mostly walk, people in a conversation gesture, some hold a cup, a phone or a folder, a few are on a phone call (pose_activity).
• Activities:
  • small groups, 4–12 people: hike (a single file or two columns along a walkable trail, facing forward), chat (2–4 in a circle), view (alone, facing a long sight line), water (0.5–3 m from a river, lake or the sea);
  • crowds, 15–40 people (tier: crowd): a tour group listening to a guide, a school outing (pupils in a two-column line and in clusters, teachers watching), picnic circles, people at a viewpoint (scenario); where a scene has too little open ground for a crowd (forest_hq_1), the episode becomes a small group. Head and chest views are rendered for 16 of them (those the exo cameras see first); everybody wears the devices and has a mask, and the cameras without a view are listed in ego_not_rendered and kept in cam_points.jsonl.
• Standing spots: slope up to 10°, flat footing, no water within 0.5 m (including mud that renders like puddles), nothing above the ground cover (trees, rocks, bushes, logs) in the body column, people at least 0.8 m apart (conversation pairs 1.1–1.5 m). Feet are set on the terrain: a planted foot at ankle height, an uphill foot sinks at most 4 cm. No two bodies overlap: each person's meshes (body, clothes, hair, devices, props) are tested against everyone within 2.2 m, and a person whose arms reach into a neighbour is rebuilt with another body and pose (or the spot stays empty; overlap_check in episode.json). Parts started after 2026-11-03 05:30 UTC also test the arms against scenery (trunks, rocks, slopes).
• Views:
  • per person, a head camera (704×704, about 100°) and a chest camera (960×540, about 110°); v2 replaces these with the device's own resolution and lens;
  • 1280×720 exo views: a pole camera (3–6 m up), a tripod (1.2–1.7 m; crowds also a far tripod) and a drone (10–25 m up). Each sees at least 2 people, and together they cover every wearer.
• Imaging: Blender Cycles at 32 samples, denoised, then a camera ISP model (<view>.png; <view>_clean.png is the render before it; parameters in isp.json).
• Horizon: Infinigen's terrain ends about 500 m out, and the sky model used to show below the horizon there. A flat far ground out to 30 km under a widened atmosphere, plus a far camera clip, closes that gap (episode.json["horizon"]). Each view records its share of sky pixels below the horizon (horizon_miss; single pixels that the atmosphere volume leaves empty in the depth pass do not count). The whole batch uses the first version of this fix (far ground at the median height of the terrain's border, 30 km): rays within about 0.4° below the horizon can still miss it, which shows as haze, mostly in the flat, hazy plain_hq_3 scene.
• Depth labels: *_depth.npy is 0 where a ray hit nothing, and also at scattered pixels of distant terrain that the atmosphere volume scattered (the RGB is unaffected; isolated holes are filled, person masks are complete).
• QA at render time (listed in dropped_views): ego views facing near scenery, blocked exo views, ego views more than 35 % covered by the wearer's own body, dark or inside-a-surface views, views with more than 0.2 % sky below the horizon, and ego views of people with fewer than 400 pixels in every exo view.
• Masks: in *_index.png, id k is person k. Every person wears both cameras, so every person is a carrier and a candidate (index_legend; non_targets is empty).
• Counts (finished parts; queries counted as in our loader: an exo view with at least 2 people of ≥ 400 px, times each ego view whose wearer is one of them; candidates = those people):

folder	scenes	episodes	people	exo / head / chest views	queries (track A)	candidates	chance
outdoor_v1/	5	146	1,775	439 / 1,358 / 1,338	6,314	11.66	0.121
outdoor_v2/	5	146	1,960	448 / 1,387 / 1,364	6,058	12.42	0.117

• Layout: outdoor_v{1,2}/<scene>_p<k>.tar holds <scene>_p<k>/summary.json and <scene>_p<k>/ep00<k>/. The file layout is that of people v1. The person records also carry activity, pose_activity, group, facing, age_group, role, ground, slope, lift, ankles and in_exo. Each episode also has tier, scenario, lens, lens_scheme and horizon.
• Camera points: outdoor_v1/cam_points.jsonl and outdoor_v2/cam_points.jsonl (schema whoego.cam_points/2, see "Camera points" below). Counts and breakdowns: outdoor_v{1,2}/stats.json.
• Generator: datagen/synth/people_nature.py (batch people_nature_batch.sh).

v0 content (v0/, outdoor_v0/; people v1 keeps the room and file layout)

• Rooms:
  • Infinigen 2 indoor rooms (living, dining, kitchen, bedroom);
  • larger living and dining rooms of 28–60 m² from our own generators, built on Infinigen;
  • work rooms, also from our own generators on Infinigen parts:
    • meeting rooms: a long table with chairs all round;
    • offices and labs: desks with office chairs and monitors along the walls, often with a small meeting table.
• People:
  • MPFB 2 bodies with the MakeHuman CC0 asset packs;
  • random adults dressed in everyday clothes;
  • standing (some face each other as if talking) or sitting on the room's own chairs and sofas;
  • up to 10 people per episode in the large rooms.
• Views per episode:
  • 1–2 exo cameras (1280×720): on the walls, framed on the group the way a human operator would;
  • per person, a head camera (704×704, about 100° horizontal FOV) and a chest camera (960×540, about 110°);
  • *_chest_lens.png: the chest view remapped to the distortion of the TF2023 chest camera (OPENCV model).
• Rendering: Blender Cycles, 64 samples, denoised. Exposure is metered on the first exo view (median brightness), with ±0.4 stop jitter.

Sets

folder	what	episodes
people_v1/	v1: BEDLAM + MPFB people, mixed facing, ISP; home rooms (big / std, tar names std_<seed>.tar) and work rooms (work)	2 per room
people_v2/	v2: people v1 with mixed real-device lenses, children, gatherings of 12–25 and everyday activities; same rooms and tar names	8 per room
v0/{big,std,work}_<seed>.tar	v0 (formerly data/): still episodes in home rooms (big / std) and work rooms (work)	2 per room
v0/v1_<n>.tar	v0 (formerly v1/): the first still-episode batch (stock home rooms, up to 6 people), same format; the fixed synthetic test set	1–2 per room
outdoor_v0/	v0 (formerly outdoor/): crowd videos: EgoGen-generated walking crowds (6–20 people) on an open square under CC0 HDRI skies and PBR ground textures, 5 s sampled at 2 fps (10 frames), 2 exo cameras on tripods / poles around the crowd, head and chest cameras of 3 wearers	1 per tar

Layout

v0/<set>_<seed>.tar            # one room; set = big | std | work (people_v1/: std_ | big_ | work_<seed>.tar)
  <set>_<seed>/summary.json    # room: free floor cells, seats, per-episode people / views / render time
  <set>_<seed>/ep000/
    episode.json               # people (id, xy, yaw, eye, sternum, height, sitting, pose, ...), views:
                               #   K, res, w2c (OpenCV), wearer id (0 for exo), visible_pixels per person id
    exo_<j>.png  exo_<j>_index.png  exo_<j>_depth.npy   # people_v1: also <view>_clean.png and isp.json
    p<k>_head.png  p<k>_head_index.png  p<k>_head_depth.npy
    p<k>_chest.png p<k>_chest_index.png p<k>_chest_depth.npy  p<k>_chest_lens*.png

Outdoor videos (outdoor_v0/open_<nn>.tar):

open_<nn>/episode.json         # fps (2), frames (source frame ids at 10 fps), setting (hdri, texture, ...), people,
                               #   wearers, views: K, res, wearer, frames: {frame id: w2c (OpenCV), visible_pixels}
open_<nn>/<view>_<frame>.png  <view>_<frame>_index.png  <view>_<frame>_depth.npy   # view = exo_<j> | p<k>_head | p<k>_chest
open_<nn>/room.json            # the open square the crowd simulation used

• *_index.png: uint8 instance mask, where 0 is background and k is person k (all of their meshes, clothes and camera props included).
• *_depth.npy: float16 z-depth in metres, 0 where there is no hit.
• The wearer of p<k>_* is person k.
• episode.json["dropped_views"]: ego views removed because the wearer's own body covered more than 70% of the frame (the view name and that share).

Camera points (cam_points.jsonl, 2026-11-02)

Where every worn / mounted camera of every episode is, in 3D and in every other view of the episode, written once by datagen/synth/cam_points.py (schema whoego.cam_points/2) so that nobody has to recompute it. Files: people_v1/cam_points.jsonl, people_v2/cam_points.jsonl, v0/cam_points.jsonl (both v0 batches; set = synth_mpfb_v2 for the {big,std,work}_<seed>.tar rooms, synth_mpfb_v1 for v1_<n>.tar) and outdoor_v0/cam_points.jsonl; episode is the folder inside the tar. Schema /2 (2026-11-02) renamed the camera list from cams to ego_in_exo, the name the real-data files use; the reverse direction of those files, exo_in_ego (exo cameras in the ego views), is not in the sim files (most sim exo views are virtual cameras with no body).

One JSON line per episode: schema, set, episode (the episode folder, the same relative path the pairs files use), flavour, tol_m (0.15), fwd_m (0.3) and ego_in_exo (the worn / mounted cameras), one entry per camera (outdoor videos: one entry per camera and frame, with frame):

• cam (camera id = its ego view name), carrier (the carrier's id in every *_index.png), mount, device, model (PINHOLE / OPENCV / OPENCV_FISHEYE), views (its own images; [] = no image: the view was dropped by QA or only some carriers have a rendered view - the device is still there and is a track A candidate), src (w2c: from the view's world-to-camera matrix; record: from the carrier record), C (camera centre, world, metres), fwd / up (unit optical axis and image-up direction, world; null when the record has no orientation), rec_off_m (distance from C to the record's camera point, which the pairs files project), off_body (only on cameras more than 1.25 body sizes from their carrier's recorded position);
• in: one entry for every other view of the episode (exo views and the other carriers' ego views; not the carrier's own): uv (pixel of C under that view's camera model, K and dist, in the pixel convention of its K; null = not projectable), z (depth along that view's axis, metres), in_frame (inside the image, and inside the image circle of circular fisheyes), d_px / id_px (rendered depth and index id at that pixel), visible (in frame and nothing more than max(0.15 m, 0.002 z) nearer than C at that pixel, or the pixel shows the carrier itself: a head camera's own head is no occluder), on_carrier, uv_fwd (pixel of C + 0.3 m along fwd, the heading in 2D).

Checks: C equals the people records' eye / sternum (0.004 px in every view; these are the candidate points of the track A queries); C + fwd lands on the principal point of the camera's own view (0.01 px).

Licence

• The images and labels are released under CC BY 4.0.
• Infinigen is BSD-3-Clause; the MakeHuman asset packs and the Poly Haven HDRIs and textures are CC0.
• MPFB is GPL-3.0. The rendered images and the generated meshes are not covered by the GPL.