Datasets:
walk_id large_stringclasses 7
values | mall_id large_stringclasses 6
values | name large_stringclasses 6
values | city large_stringclasses 6
values | state large_stringclasses 3
values | contributor large_stringclasses 6
values | date_local large_stringdate 2026-09-19 00:00:00 2026-09-20 00:00:00 | center large_stringclasses 7
values | orientation large_stringclasses 2
values | zoom float64 0.5 1 | note large_stringclasses 2
values | clips int64 156 645 | videos int64 156 498 | photos int64 0 147 | video_seconds float64 737 2.42k | clip_duration_s_med float64 4.88 5.03 | first_capture_utc large_stringdate 2026-09-19 15:21:50 2026-09-20 22:10:32 | last_capture_utc large_stringdate 2026-09-19 17:02:08 2026-09-21 00:06:32 | span_minutes float64 48.2 168 | device_model large_stringclasses 5
values | os_version large_stringclasses 4
values | app_version large_stringclasses 1
value | extract_fps float64 4 4 | gps_quality large_stringclasses 2
values | gps_fixes int64 156 645 | gps_accuracy_med_m float64 11.8 1.82k | gps_hull_area_m2 float64 21.5k 28.1k ⌀ | altitude_med_m float64 0 169 | frames_posed int64 3.1k 9.94k | frames_extracted int64 3.1k 10.1k | path_m float64 1.06k 6.02k | extent_max_m float64 67.4 605 | extent_p95_m float64 58.1 521 | vertical_extent_m float64 5.6 50.4 | depth_m_med float64 2.78 7.95 | depth_m_p10 float64 1.86 6.63 | depth_m_p90 float64 3.73 15.5 | depth_valid_frac_med float64 0.82 0.93 | overlap_next_med float64 0.75 0.87 | overlap_next_p10 float64 0.32 0.79 | parallax_deg_med float64 1.23 3.96 | n_covisible_med float64 11 16 | revisits_found float64 1 9 ⌀ | loops_closed float64 1 5 ⌀ | loop_resid_rms_m float64 0.24 2.57 ⌀ | gps_prior bool 1
class | gps_resid_med_m float64 10.7 26.2 ⌀ | gps_global_scale float64 0.87 1.45 ⌀ | gps_yaw_deg float64 102 174 ⌀ | media_pack null |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
champlain_centre | champlain_centre | Champlain Centre | Plattsburgh | NY | contributor_1 | 2026-09-19 | [44.702319, -73.491183] | portrait | 1 | null | 227 | 227 | 0 | 807.8 | 5.03 | 2026-09-19T15:21:50+00:00 | 2026-09-19T17:02:08+00:00 | 100.3 | iPhone14,4 | 26.6.2 | 1.1.17 | 4 | precise | 227 | 11.8 | 25,610 | 98 | 4,456 | 4,456 | 1,715.8 | 146.6 | 131.9 | 17.2 | 5.39 | 3.25 | 6.94 | 0.913 | 0.829 | 0.552 | 1.859 | 12 | 9 | 5 | 0.4945 | true | 16.5 | 0.8659 | 102.4 | null |
university_mall | university_mall | University Mall | South Burlington | VT | contributor_1 | 2026-09-20 | [44.465449, -73.183937] | portrait | 0.5 | mixed 0.5x / 1x zoom during the walk, reconstructed with one walk-wide K | 156 | 156 | 0 | 736.8 | 5 | 2026-09-20T16:55:57+00:00 | 2026-09-20T17:44:12+00:00 | 48.2 | iPhone14,4 | 26.6.2 | 1.1.17 | 4 | precise | 156 | 23.8 | 28,140 | 94.3 | 3,101 | 3,101 | 1,341.5 | 127.5 | 120.8 | 50.4 | 2.78 | 1.86 | 3.73 | 0.896 | 0.754 | 0.4 | 3.961 | 11 | 8 | 2 | 2.5718 | true | 14.8 | 1.0323 | 101.8 | null |
atlantic_terminal | atlantic_terminal | Atlantic Terminal Mall | Brooklyn | NY | contributor_2 | 2026-09-20 | [40.684114, -73.977129] | portrait | 1 | null | 283 | 210 | 73 | 1,050 | 5.03 | 2026-09-20T17:18:38+00:00 | 2026-09-20T19:07:35+00:00 | 109 | iPhone18,3 | 26.6.1 | 1.1.17 | 4 | approximate | 283 | 1,414 | null | 0 | 4,178 | 4,251 | 2,415.2 | 67.4 | 58.1 | 20.8 | 6.66 | 3.8 | 9.38 | 0.93 | 0.813 | 0.321 | 2.253 | 11 | 8 | 3 | 0.5868 | null | null | null | null | null |
glendale_galleria_a | glendale_galleria | Glendale Galleria | Glendale | CA | contributor_3 | 2026-09-20 | [34.145148, -118.259197] | portrait | 1 | loop closure pending: poses are the chained first pass, expect drift on long corridors | 645 | 498 | 147 | 2,416.1 | 5 | 2026-09-20T20:11:42+00:00 | 2026-09-20T22:57:49+00:00 | 166.1 | iPhone18,2 | 26.6.2 | 1.1.17 | 4 | approximate | 645 | 1,414 | null | 161.8 | 9,943 | 10,090 | 6,022.7 | 605.1 | 521 | 16.4 | 6.56 | 5.16 | 11.27 | 0.884 | 0.84 | 0.617 | 1.969 | 14 | null | null | null | null | null | null | null | null |
glendale_galleria_b | glendale_galleria | Glendale Galleria | Glendale | CA | contributor_4 | 2026-09-20 | [34.145125, -118.258487] | landscape | 1 | loop closure pending: poses are the chained first pass, expect drift on long corridors | 536 | 396 | 140 | 1,940.3 | 4.88 | 2026-09-20T20:11:20+00:00 | 2026-09-20T22:59:19+00:00 | 168 | iPhone19,3 | 27.0 | 1.1.17 | 4 | approximate | 532 | 1,823.1 | null | 168.7 | 7,892 | 8,032 | 5,819.4 | 218.6 | 172.6 | 25.9 | 6.95 | 4.72 | 9.28 | 0.838 | 0.805 | 0.37 | 1.926 | 13 | null | null | null | null | null | null | null | null |
westfield_north_county | westfield_north_county | Westfield North County | Escondido | CA | contributor_5 | 2026-09-20 | [33.071064, -117.065806] | landscape | 1 | null | 333 | 206 | 127 | 1,026.5 | 5 | 2026-09-20T22:10:32+00:00 | 2026-09-21T00:06:32+00:00 | 116 | iPhone16,1 | 26.6.2 | 1.1.17 | 4 | precise | 333 | 19.7 | 21,495 | 109.6 | 4,110 | 4,237 | 1,058.3 | 162.7 | 157.3 | 5.6 | 7.17 | 5.43 | 9.71 | 0.815 | 0.865 | 0.785 | 1.228 | 16 | 1 | 1 | 0.2395 | true | 10.7 | 1.0469 | 127.1 | null |
hudson_yards | hudson_yards | The Shops at Hudson Yards | New York | NY | contributor_6 | 2026-09-20 | [40.753617, -74.001426] | portrait | 1 | null | 369 | 301 | 68 | 1,499.7 | 5 | 2026-09-20T21:11:24+00:00 | 2026-09-20T22:59:58+00:00 | 108.6 | iPhone18,2 | 26.6 | 1.1.17 | 4 | precise | 369 | 13.9 | 23,210 | 17.2 | 6,020 | 6,088 | 2,134.1 | 80.8 | 70.5 | 14.3 | 7.95 | 6.63 | 15.53 | 0.89 | 0.871 | 0.674 | 1.642 | 16 | 5 | 4 | 1.1837 | true | 26.2 | 1.4485 | 173.8 | null |
US Malls (sample): indoor mall walks with camera poses and metric depth
v0.1 · built 2026-10-08 · collectivememory/us-malls-public
This is a sample of the Collective Memory mall collection: 7 walks through 6 US malls, de-identified. The full collection has more malls and cities, high-resolution media and IMU data. For access, contact info@collectivememory.ai.
Contributors walked each mall end to end with the Collective Memory app, recording back-to-back 5-second clips along every
corridor in both directions. Every frame comes with a camera pose and metric depth, every clip with its GPS fix and capture
metadata, and a relocalization benchmark is included (2549 clips, 39700 posed frames). Faces and licence
plates are blurred in every video, photo and frame; clip ids are salted hashes and contributors are labelled contributor_N.
 Champlain Centre Plattsburgh, NY · 4,456 frames · 1,716 m |  University Mall South Burlington, VT · 3,101 frames · 1,342 m |  Westfield North County Escondido, CA · 4,110 frames · 1,058 m |  The Shops at Hudson Yards New York, NY · 6,020 frames · 2,134 m |
| GPS was less accurate on these walks, as is common deep inside large or multi-level buildings, so the track is placed on the map approximately | |||
 Atlantic Terminal Mall Brooklyn, NY · 4,178 frames · 2,415 m |  Glendale Galleria Glendale, CA · 9,943 frames · 6,023 m |  Glendale Galleria Glendale, CA · 7,892 frames · 5,819 m | |
At a glance
| malls | 6 · Glendale Galleria (2 observers) |
| walks | 7 |
| where | 6 cities in 3 states: Brooklyn, NY; Escondido, CA; Glendale, CA; New York, NY; Plattsburgh, NY; South Burlington, VT |
| when | 2026-09-19 to 2026-09-21 (UTC) |
| media | 1,994 videos (158 min, ~5 s each) + 555 photos, faces and licence plates blurred |
| frames | 39,700 video frames at 4 fps, all posed with metric depth; 40,255 shard samples incl. the photos |
| walked | 20.5 km of reconstructed camera path, 13.6 h of capture sessions |
| size | 16.7 GB: media 4.5 GB, frames + depth shards 11.8 GB, tables + maps 0.04 GB |
The wider collection
Collective Memory has 18.6M clips (15.5M photos, 3.0M videos) from 172,792 contributors in 174 countries (through 2026-09); 102,863 mappers uploaded in the past 12 months. In the US, 2,342 contributors have geolocated uploads across 50 states and DC; the most in CA 522 · NY 412 · FL 295 · DC 230 · TX 148 · NJ 112 · MA 97 · PA 78 · WA 75 · IL 74.
These walks are a sample. For more malls, high-resolution media and IMU data, contact info@collectivememory.ai.
What's inside
- Videos and photos: every clip of every walk (~5 s each), faces and plates blurred, H.264
.mp4without audio or metadata. - Posed frames: 4 fps frames, each with a camera-to-world pose in metres and a metric depth map (MapAnything), loop-closed and, where the phone had precise GPS, aligned to the map. Packed as WebDataset shards.
- Capture metadata: per clip GPS fix (accuracy, altitude), time, device, OS and app version.
- Maps per walk: track on satellite imagery, coverage map, sample frames with depth, timed GeoJSON trail.
- Relocalization benchmark (
localization/): relocalize a frame from one visit of a place against a map from another visit; 18 places, 784 scored queries, hloc baseline 58% within 10 cm, 1°.
Poses and depth are a first, feed-forward pass (no bundle adjustment yet): good for coverage, sampling and depth statistics;
treat the 3-D poses as preliminary. Media paths and clip_ids are stable across revisions.
Malls
| mall | location | date | clips | video (min) | posed frames | path (m) | median depth (m) | GPS |
|---|---|---|---|---|---|---|---|---|
| Champlain Centre | Plattsburgh, NY | 2026-09-19 | 227 | 13.5 | 4,456 | 1,716 | 5.4 | precise |
| University Mall | South Burlington, VT | 2026-09-20 | 156 | 12.3 | 3,101 | 1,342 | 2.8 | precise |
| Atlantic Terminal Mall | Brooklyn, NY | 2026-09-20 | 283 | 17.5 | 4,178 | 2,415 | 6.7 | approximate |
| Glendale Galleria | Glendale, CA | 2026-09-20 | 645 | 40.3 | 9,943 | 6,023 | 6.6 | approximate |
| Glendale Galleria | Glendale, CA | 2026-09-20 | 536 | 32.3 | 7,892 | 5,819 | 7.0 | approximate |
| Westfield North County | Escondido, CA | 2026-09-20 | 333 | 17.1 | 4,110 | 1,058 | 7.2 | precise |
| The Shops at Hudson Yards | New York, NY | 2026-09-20 | 369 | 25.0 | 6,020 | 2,134 | 8.0 | precise |
Download
from huggingface_hub import snapshot_download
snapshot_download("collectivememory/us-malls-public", repo_type="dataset") # everything
snapshot_download("collectivememory/us-malls-public", repo_type="dataset", allow_patterns=["<walk_id>/**", "malls.parquet"]) # one mall
snapshot_download("collectivememory/us-malls-public", repo_type="dataset", allow_patterns=["*.parquet", "*/*.parquet", "*/mall.json", "*/maps/*"]) # tables + figures
snapshot_download("collectivememory/us-malls-public", repo_type="dataset", allow_patterns=["localization/**"]) # relocalization benchmark
Layout and file formats
malls.parquet one row per walk: mall, location, contributor, date, counts, span, path length, depth / overlap stats, device, lens
clips.parquet every clip of every walk (see <walk_id>/clips.parquet)
frames.parquet every posed frame of every walk (see <walk_id>/frames.parquet)
localization/ relocalization benchmark: per place images, metric COLMAP reference, splits, baseline, eval.py (its README)
assets/global/ the coverage map above
<walk_id>/
mall.json stats (capture, coverage, depth, GPS) + reconstruction settings and QA for this walk
clips.parquet one row per clip: clip_id, created (UTC), media, duration_s, width / height / mime_type,
lat / lng / gps_accuracy_m / gps_quality / altitude_m, device_model / os_version / app_version,
orientation, zoom, media_path, n_frames / n_frames_posed, fps, frame_shard
frames.parquet one row per posed frame: frame_key (<clip_id>__fNNN), walk_frame_idx, t_rel_s, image and depth size,
fx fy cx cy (at depth size) and fx_img.. (at image size), c2w (16 floats, OpenCV cam->world, metres),
median_depth_m, valid_frac, depth_conf, parallax_deg, overlap_prev / overlap_next, n_covisible,
window_id, window_resid_m, frame_shard
envelopes.parquet the full flattened capture envelope per clip (location_*, media_*, device_*, time_*)
maps/ walk.png (track), overlap.png (coverage), examples.png (frames + depth), walk.geojson (timed trail)
videos/<clip_id>.mp4 540x960 (or 960x540), faces and plates blurred, H.264, no audio or metadata
photos/<clip_id>.jpg long side <= 1920 px, faces and plates blurred
frames/shard-NNN.tar WebDataset, 1000 samples per shard, capture order: <frame_key>.jpg (4 fps, <= 30 per clip, blurred)
+ <frame_key>.npz when posed: depth (float16, metres, model output size), conf (uint8), mask,
K (3x3), c2w (4x4). A photo is one sample, <clip_id>__f001.jpg, with no .npz.
How it was made
Capture. Contributors "paint" a mall: walk the full perimeter of the public walkway recording back-to-back 5-second clips with the phone pointed ahead and gently panning, then the same route in reverse, then the connecting corridors. Stores are not entered. One contributor, one mall, one session (1-3 h); Glendale Galleria was walked by two contributors at the same time. The app uploads each clip with a capture envelope (GPS fix with accuracy and altitude, media size and duration, device, OS, app version, capture time); video reaches the server at 540x960.
Reconstruction. Frames at 4 fps; poses and metric depth from MapAnything over sliding windows of 64 frames, chained
into one metric frame; revisits found with DINOv3 + MegaLoc and closed in a pose graph, which on precise-GPS walks also
carries each clip's GPS fix. Depth is at the model's output size (long side 518); intrinsics are given at both sizes.
mall.json records the settings and QA per walk.
De-identification. Faces (YOLOv8-face + YuNet) and licence plates (YOLO11) are detected on every frame, photo and benchmark image and blurred; videos are processed frame by frame with detections linked into tracks, so a face missed for a frame stays blurred, then re-encoded without audio or metadata. The detector errs on the side of blurring: some signs (EXIT signs, logos) are blurred too.
Relocalization benchmark
Relocalization benchmark
localization/: pose of a frame from visit B of a place given a map from visit A (same contributor, minutes earlier), both
directions; 18 places, 7 walks, 540x960. Pseudo ground truth: two independent metric SfM references (COLMAP
SIFT; DISK + LightGlue); scored queries: references agree <= 10 cm, <= 1°. Baseline: hloc (NetVLAD top-10, ALIKED + LightGlue,
PnP-RANSAC 12 px); unlocalized = failure.
| mall | phone | capture | places scored | query frames | within 10 cm, 1° | within 25 cm, 2° | within 1 m, 5° | median error | not scored |
|---|---|---|---|---|---|---|---|---|---|
| Champlain Centre | iPhone14,4 | 540x960, 1.0x | 2 of 3 | 217 | 36% | 50% | 57% | 26 cm | 1 place: references disagree |
| University Mall | iPhone14,4 | 540x960, 0.5x + 1.0x | 0 of 3 | 0 | – | – | – | – | mixed 0.5x / 1x zoom; 3 places: references disagree |
| Atlantic Terminal Mall | iPhone18,3 | 540x960, 1.0x | 1 of 3 | 64 | 16% | 56% | 62% | 21 cm | 2 places: visits not co-registered |
| Glendale Galleria (walk A) | iPhone18,2 | 540x960, 1.0x | 3 of 3 | 255 | 93% | 95% | 96% | 1.0 cm | |
| Glendale Galleria (walk B) | iPhone19,3 | 960x540, 1.0x | 2 of 3 | 66 | 58% | 67% | 70% | 7.4 cm | 1 place: references disagree |
| Westfield North County | iPhone16,1 | 960x540, 1.0x | 1 of 1 | 100 | 51% | 64% | 67% | 8.3 cm | |
| The Shops at Hudson Yards | iPhone18,2 | 540x960, 1.0x | 1 of 2 | 82 | 52% | 55% | 56% | 7.9 cm | 1 place: references disagree |
| all scored | 784 | 58% | 69% | 72% | 5.9 cm | ||||
| same way (< 45°) | 630 | 65% | 75% | 77% | 4.7 cm | ||||
| wide angle (> 45°) | 154 | 32% | 45% | 53% | 39 cm |
Capture settings supported by these results:
| setting | measured | rule |
|---|---|---|
| zoom | 1x, fixed: 10 of 15 places scored (6 walks); 0.5x mixed with 1x: 0 of 3 places scored (1 walk) | lock 1x |
| return direction | within 10 cm, 1°: 65% same way (< 45°, 630 queries), 32% wide angle (> 45°, 154 queries) | walk every corridor in both directions |
| orientation | same-way returns within 10 cm, 1°: portrait 65% (516 queries, 4 walks), landscape 63% (114 queries, 2 walks); no measured difference. Mixed within a walk: atlantic_terminal: 27 of 210 clips in the other orientation, 1 of 3 places scored (1 walk) | either; one orientation per walk |
Protocol, formats, evaluator: localization/README.md.
Per-place and per-direction numbers: localization/DETAILS.md.
Walks
Champlain Centre · champlain_centre
Plattsburgh, NY · 44.70232, -73.49118 · walked 2026-09-19 by contributor_1 on iPhone14,4 (OS 26.6.2), portrait 1.0x · 227 clips, 13 min of video over 100 min
Track: time-gradient trail start → end on satellite imagery |  Coverage: trajectory coloured by frame-to-frame overlap, per-frame overlap / parallax strip |
 Examples: frames along the walk with their metric depth | |
13:39 into the walk · 5.0 s · m_748e889acd3c950f.mp4 | 31:49 into the walk · 5.0 s · m_801bc3505f7b3dbd.mp4 | 50:46 into the walk · 5.0 s · m_f392d6e05ba07907.mp4 | 82:11 into the walk · 5.0 s · m_d45f85894f4cbc5c.mp4 |
De-identified clips (faces and plates blurred, re-encoded, no audio) (540x960), 227 in this walk under champlain_centre/videos/.
| Capture | Coverage | ||
| clips (video / photo) | 227 (227 / 0) | video frames posed @ fps / shard samples | 4,456 @ 4 fps / 4,456 |
| video | 13.5 min | path length | 1,715.8 m |
| median clip | 5.03 s | footprint (p95 / max radius) | 131.9 m / 146.6 m |
| first → last upload (UTC) | 15:21:50 → 17:02:08 | vertical extent | 17.2 m |
| capture / uploaded size | 1080x1920 / 540x960 | GPS hull area | 25,610 m² |
| phone / OS / app | iPhone14,4 / 26.6.2 / 1.1.17 | revisits found / loops closed | 9 / 5 (RMS 0.49 m) |
| Depth and multi-view | GPS and georeferencing | ||
| median depth (p10 / p90) | 5.4 m (3.2 m / 6.9 m) | GPS quality | precise, median accuracy 11.8 m |
| valid depth per frame | 0.91 | fixes | 227 |
| overlap to next frame (median / p10) | 0.83 / 0.55 | altitude (median) | 98.0 m |
| parallax to next frame | 1.86° | reconstruction | pose graph with GPS edges, median residual 16.5 m, scale x0.87 |
| covisible frames | 12 | ||
| depth map size / FOV | 294x518 · 30° x 51° |
champlain_centre/clips.parquet (227 rows):
| clip_id | created | media | duration_s | width | height | lat | lng | gps_accuracy_m | altitude_m | n_frames | media_path |
|---|---|---|---|---|---|---|---|---|---|---|---|
| m_26356491bca1ca78 | 2026-09-19 15:21:50Z | video | 5.031 | 1,080 | 1,920 | 44.7034 | -73.4937 | 4.6 | 97.6 | 20 | champlain_centre/videos/m_26356491bca1ca78.mp4 |
| m_2e4ba46c609a8d56 | 2026-09-19 15:22:13Z | video | 5.031 | 1,080 | 1,920 | 44.7034 | -73.4935 | 3 | 98.7 | 20 | champlain_centre/videos/m_2e4ba46c609a8d56.mp4 |
| m_d5bd8fd69cf245dc | 2026-09-19 15:23:06Z | video | 5.031 | 1,080 | 1,920 | 44.7035 | -73.4931 | 20.4 | 97 | 20 | champlain_centre/videos/m_d5bd8fd69cf245dc.mp4 |
champlain_centre/frames.parquet (4,456 rows):
| frame_key | walk_frame_idx | t_rel_s | median_depth_m | valid_frac | overlap_next | parallax_deg | n_covisible | window_id | c2w |
|---|---|---|---|---|---|---|---|---|---|
| m_26356491bca1ca78__f001 | 0 | 0 | 3.21178 | 0.551808 | 0.81 | 2.597 | 8 | 0 | [16 floats] |
| m_26356491bca1ca78__f002 | 1 | 0.25 | 3.227 | 0.554927 | 0.811 | 3.372 | 9 | 0 | [16 floats] |
| m_26356491bca1ca78__f003 | 2 | 0.5 | 3.33525 | 0.556444 | 0.803 | 4.176 | 10 | 0 | [16 floats] |
University Mall · university_mall
South Burlington, VT · 44.46545, -73.18394 · walked 2026-09-20 by contributor_1 on iPhone14,4 (OS 26.6.2), portrait 0.5x · 156 clips, 12 min of video over 48 min
Caveat: mixed 0.5x / 1x zoom during the walk, reconstructed with one walk-wide K.
Track: time-gradient trail start → end on satellite imagery |  Coverage: trajectory coloured by frame-to-frame overlap, per-frame overlap / parallax strip |
 Examples: frames along the walk with their metric depth | |
04:29 into the walk · 5.0 s · m_a8c013bbbe220ca2.mp4 | 13:32 into the walk · 5.0 s · m_97fb55e6a3804ed9.mp4 | 24:50 into the walk · 5.0 s · m_17107a1a9f850f1e.mp4 | 43:46 into the walk · 5.0 s · m_dac8f5a1d4fbf107.mp4 |
De-identified clips (faces and plates blurred, re-encoded, no audio) (540x960), 156 in this walk under university_mall/videos/.
| Capture | Coverage | ||
| clips (video / photo) | 156 (156 / 0) | video frames posed @ fps / shard samples | 3,101 @ 4 fps / 3,101 |
| video | 12.3 min | path length | 1,341.5 m |
| median clip | 5.00 s | footprint (p95 / max radius) | 120.8 m / 127.5 m |
| first → last upload (UTC) | 16:55:57 → 17:44:12 | vertical extent | 50.4 m |
| capture / uploaded size | 1080x1920 / 540x960 | GPS hull area | 28,140 m² |
| phone / OS / app | iPhone14,4 / 26.6.2 / 1.1.17 | revisits found / loops closed | 8 / 2 (RMS 2.57 m) |
| Depth and multi-view | GPS and georeferencing | ||
| median depth (p10 / p90) | 2.8 m (1.9 m / 3.7 m) | GPS quality | precise, median accuracy 23.8 m |
| valid depth per frame | 0.90 | fixes | 156 |
| overlap to next frame (median / p10) | 0.75 / 0.40 | altitude (median) | 94.3 m |
| parallax to next frame | 3.96° | reconstruction | pose graph with GPS edges, median residual 14.8 m, scale x1.03 |
| covisible frames | 11 | ||
| depth map size / FOV | 294x518 · 57° x 88° |
university_mall/clips.parquet (156 rows):
| clip_id | created | media | duration_s | width | height | lat | lng | gps_accuracy_m | altitude_m | n_frames | media_path |
|---|---|---|---|---|---|---|---|---|---|---|---|
| m_007c9e6eed210357 | 2026-09-20 16:55:57Z | video | 5.001 | 1,080 | 1,920 | 44.466 | -73.183 | 27 | 94.3 | 20 | university_mall/videos/m_007c9e6eed210357.mp4 |
| m_ee2212ef17145264 | 2026-09-20 16:56:13Z | video | 5.031 | 1,080 | 1,920 | 44.4659 | -73.1833 | 31.7 | 94.2 | 20 | university_mall/videos/m_ee2212ef17145264.mp4 |
| m_1e8103071d322102 | 2026-09-20 16:56:24Z | video | 5.031 | 1,080 | 1,920 | 44.466 | -73.1833 | 23.2 | 94.2 | 20 | university_mall/videos/m_1e8103071d322102.mp4 |
university_mall/frames.parquet (3,101 rows):
| frame_key | walk_frame_idx | t_rel_s | median_depth_m | valid_frac | overlap_next | parallax_deg | n_covisible | window_id | c2w |
|---|---|---|---|---|---|---|---|---|---|
| m_007c9e6eed210357__f001 | 0 | 0 | 4.25811 | 0.926923 | 0.965 | 0.44 | 8 | 0 | [16 floats] |
| m_007c9e6eed210357__f002 | 1 | 0.25 | 4.22585 | 0.924165 | 0.951 | 0.192 | 9 | 0 | [16 floats] |
| m_007c9e6eed210357__f003 | 2 | 0.5 | 4.25408 | 0.926418 | 0.929 | 0.562 | 10 | 0 | [16 floats] |
Atlantic Terminal Mall · atlantic_terminal
Brooklyn, NY · 40.68411, -73.97713 · walked 2026-09-20 by contributor_2 on iPhone18,3 (OS 26.6.1), portrait 1.0x · 283 clips, 18 min of video over 109 min
Track: time-gradient trail start → end on satellite imagery |  Coverage: trajectory coloured by frame-to-frame overlap, per-frame overlap / parallax strip |
 Examples: frames along the walk with their metric depth | |
10:47 into the walk · 5.0 s · m_8f9837accb3f7e0a.mp4 | 25:59 into the walk · 5.0 s · m_f6d1d1deb7c8ac56.mp4 | 58:45 into the walk · 5.0 s · m_fb3c616bc3f78509.mp4 | 91:04 into the walk · 5.0 s · m_2d6d8872eadfa494.mp4 |
De-identified clips (faces and plates blurred, re-encoded, no audio) (540x960), 210 in this walk under atlantic_terminal/videos/.
| Capture | Coverage | ||
| clips (video / photo) | 283 (210 / 73) | video frames posed @ fps / shard samples | 4,178 @ 4 fps / 4,251 (incl. 73 photos) |
| video | 17.5 min | path length | 2,415.2 m |
| median clip | 5.03 s | footprint (p95 / max radius) | 58.1 m / 67.4 m |
| first → last upload (UTC) | 17:18:38 → 19:07:35 | vertical extent | 20.8 m |
| capture / uploaded size | 1080x1920 / 540x960 | GPS hull area | - |
| phone / OS / app | iPhone18,3 / 26.6.1 / 1.1.17 | revisits found / loops closed | 8 / 3 (RMS 0.59 m) |
| Depth and multi-view | GPS and georeferencing | ||
| median depth (p10 / p90) | 6.7 m (3.8 m / 9.4 m) | GPS quality | approximate, median accuracy 1,414.0 m |
| valid depth per frame | 0.93 | fixes | 283 |
| overlap to next frame (median / p10) | 0.81 / 0.32 | altitude (median) | 0.0 m |
| parallax to next frame | 2.25° | reconstruction | not georeferenced (approximate location only) |
| covisible frames | 11 | ||
| depth map size / FOV | 294x518 · 30° x 51° |
atlantic_terminal/clips.parquet (283 rows):
| clip_id | created | media | duration_s | width | height | lat | lng | gps_accuracy_m | altitude_m | n_frames | media_path |
|---|---|---|---|---|---|---|---|---|---|---|---|
| m_a68261c7abf8295f | 2026-09-20 17:18:38Z | video | 5.011 | 1080 | 1920 | 40.684 | -73.9771 | 1,389 | 0 | 20 | atlantic_terminal/videos/m_a68261c7abf8295f.mp4 |
| m_52612d8e126d0438 | 2026-09-20 17:19:37Z | video | 5.013 | 1080 | 1920 | 40.6852 | -73.9778 | 1,414 | 0 | 20 | atlantic_terminal/videos/m_52612d8e126d0438.mp4 |
| m_ef6294d7cfb7d0a2 | 2026-09-20 17:20:00Z | video | 5.011 | 1080 | 1920 | 40.6837 | -73.9774 | 1,414 | 0 | 20 | atlantic_terminal/videos/m_ef6294d7cfb7d0a2.mp4 |
atlantic_terminal/frames.parquet (4,178 rows):
| frame_key | walk_frame_idx | t_rel_s | median_depth_m | valid_frac | overlap_next | parallax_deg | n_covisible | window_id | c2w |
|---|---|---|---|---|---|---|---|---|---|
| m_a68261c7abf8295f__f001 | 0 | 0 | 6.26953 | 0.942282 | 0.903 | 1.308 | 8 | 0 | [16 floats] |
| m_a68261c7abf8295f__f002 | 1 | 0.25 | 6.21094 | 0.940424 | 0.92 | 1.282 | 9 | 0 | [16 floats] |
| m_a68261c7abf8295f__f003 | 2 | 0.5 | 6.17578 | 0.938447 | 0.879 | 1.987 | 10 | 0 | [16 floats] |
Glendale Galleria · glendale_galleria_a
Glendale, CA · 34.14515, -118.25920 · walked 2026-09-20 by contributor_3 on iPhone18,2 (OS 26.6.2), portrait 1.0x · 645 clips, 40 min of video over 166 min
Caveat: loop closure pending: poses are the chained first pass, expect drift on long corridors.
Track: time-gradient trail start → end on satellite imagery |  Coverage: trajectory coloured by frame-to-frame overlap, per-frame overlap / parallax strip |
 Examples: frames along the walk with their metric depth | |
22:53 into the walk · 5.0 s · m_b9dfd8ff3722f72c.mp4 | 58:56 into the walk · 5.0 s · m_f90ce06658f3ed78.mp4 | 97:35 into the walk · 5.0 s · m_9d0b396f9b449ca1.mp4 | 138:02 into the walk · 5.0 s · m_a72312c602539fe8.mp4 |
De-identified clips (faces and plates blurred, re-encoded, no audio) (540x960), 498 in this walk under glendale_galleria_a/videos/.
| Capture | Coverage | ||
| clips (video / photo) | 645 (498 / 147) | video frames posed @ fps / shard samples | 9,943 @ 4 fps / 10,090 (incl. 147 photos) |
| video | 40.3 min | path length | 6,022.7 m |
| median clip | 5.00 s | footprint (p95 / max radius) | 521.0 m / 605.1 m |
| first → last upload (UTC) | 20:11:42 → 22:57:49 | vertical extent | 16.4 m |
| capture / uploaded size | 1080x1920 / 540x960 | GPS hull area | - |
| phone / OS / app | iPhone18,2 / 26.6.2 / 1.1.17 | revisits found / loops closed | - / - |
| Depth and multi-view | GPS and georeferencing | ||
| median depth (p10 / p90) | 6.6 m (5.2 m / 11.3 m) | GPS quality | approximate, median accuracy 1,414.0 m |
| valid depth per frame | 0.88 | fixes | 645 |
| overlap to next frame (median / p10) | 0.84 / 0.62 | altitude (median) | 161.8 m |
| parallax to next frame | 1.97° | reconstruction | not georeferenced (approximate location only) |
| covisible frames | 14 | ||
| depth map size / FOV | 294x518 · 31° x 52° |
glendale_galleria_a/clips.parquet (645 rows):
| clip_id | created | media | duration_s | width | height | lat | lng | gps_accuracy_m | altitude_m | n_frames | media_path |
|---|---|---|---|---|---|---|---|---|---|---|---|
| m_2df9a9053b8aeadc | 2026-09-20 20:11:42Z | video | 4.97 | 1,080 | 1,920 | 34.1444 | -118.259 | 194.6 | 157.9 | 20 | glendale_galleria_a/videos/m_2df9a9053b8aeadc.mp4 |
| m_2cb28e34b9ca44bf | 2026-09-20 20:12:00Z | video | 4.948 | 1,080 | 1,920 | 34.1444 | -118.259 | 439.5 | 158.1 | 20 | glendale_galleria_a/videos/m_2cb28e34b9ca44bf.mp4 |
| m_700df8498a23099d | 2026-09-20 20:12:22Z | video | 5 | 1,080 | 1,920 | 34.1443 | -118.259 | 667 | 158.1 | 20 | glendale_galleria_a/videos/m_700df8498a23099d.mp4 |
glendale_galleria_a/frames.parquet (9,943 rows):
| frame_key | walk_frame_idx | t_rel_s | median_depth_m | valid_frac | overlap_next | parallax_deg | n_covisible | window_id | c2w |
|---|---|---|---|---|---|---|---|---|---|
| m_2df9a9053b8aeadc__f001 | 0 | 0 | 13.7578 | 0.897165 | 0.895 | 0.932 | 8 | 0 | [16 floats] |
| m_2df9a9053b8aeadc__f002 | 1 | 0.25 | 13.5781 | 0.89855 | 0.935 | 0.525 | 9 | 0 | [16 floats] |
| m_2df9a9053b8aeadc__f003 | 2 | 0.5 | 13.4922 | 0.897939 | 0.903 | 0.935 | 10 | 0 | [16 floats] |
Glendale Galleria · glendale_galleria_b
Glendale, CA · 34.14513, -118.25849 · walked 2026-09-20 by contributor_4 on iPhone19,3 (OS 27.0), landscape 1.0x · 536 clips, 32 min of video over 168 min
Caveat: loop closure pending: poses are the chained first pass, expect drift on long corridors.
Track: time-gradient trail start → end on satellite imagery |  Coverage: trajectory coloured by frame-to-frame overlap, per-frame overlap / parallax strip |
 Examples: frames along the walk with their metric depth | |
18:19 into the walk · 4.9 s · m_536cd1329c82eb46.mp4 | 57:36 into the walk · 4.9 s · m_081e8e90a04c01ca.mp4 | 97:17 into the walk · 5.0 s · m_2506abd1bbf668c0.mp4 | 147:54 into the walk · 5.0 s · m_817d9c65e74971af.mp4 |
De-identified clips (faces and plates blurred, re-encoded, no audio) (960x540), 396 in this walk under glendale_galleria_b/videos/.
| Capture | Coverage | ||
| clips (video / photo) | 536 (396 / 140) | video frames posed @ fps / shard samples | 7,892 @ 4 fps / 8,032 (incl. 140 photos) |
| video | 32.3 min | path length | 5,819.4 m |
| median clip | 4.88 s | footprint (p95 / max radius) | 172.6 m / 218.6 m |
| first → last upload (UTC) | 20:11:20 → 22:59:19 | vertical extent | 25.9 m |
| capture / uploaded size | 1920x1080 / 960x540 | GPS hull area | - |
| phone / OS / app | iPhone19,3 / 27.0 / 1.1.17 | revisits found / loops closed | - / - |
| Depth and multi-view | GPS and georeferencing | ||
| median depth (p10 / p90) | 7.0 m (4.7 m / 9.3 m) | GPS quality | approximate, median accuracy 1,823.1 m |
| valid depth per frame | 0.84 | fixes | 532 |
| overlap to next frame (median / p10) | 0.81 / 0.37 | altitude (median) | 168.7 m |
| parallax to next frame | 1.93° | reconstruction | not georeferenced (approximate location only) |
| covisible frames | 13 | ||
| depth map size / FOV | 518x294 · 62° x 38° |
glendale_galleria_b/clips.parquet (536 rows):
| clip_id | created | media | duration_s | width | height | lat | lng | gps_accuracy_m | altitude_m | n_frames | media_path |
|---|---|---|---|---|---|---|---|---|---|---|---|
| m_d411f369ab593053 | 2026-09-20 20:11:20Z | video | 4.883 | 1920 | 1080 | 34.1444 | -118.259 | 14.8 | 156.5 | 20 | glendale_galleria_b/videos/m_d411f369ab593053.mp4 |
| m_af41b4c05dc32c01 | 2026-09-20 20:11:34Z | video | 4.883 | 1920 | 1080 | 34.1442 | -118.259 | 18.7 | 157.5 | 20 | glendale_galleria_b/videos/m_af41b4c05dc32c01.mp4 |
| m_1e58c18e3d655048 | 2026-09-20 20:11:53Z | video | 4.863 | 1920 | 1080 | 34.1439 | -118.259 | 4.6 | 157.5 | 19 | glendale_galleria_b/videos/m_1e58c18e3d655048.mp4 |
glendale_galleria_b/frames.parquet (7,892 rows):
| frame_key | walk_frame_idx | t_rel_s | median_depth_m | valid_frac | overlap_next | parallax_deg | n_covisible | window_id | c2w |
|---|---|---|---|---|---|---|---|---|---|
| m_d411f369ab593053__f001 | 0 | 0 | 5.90625 | 0.8163 | 0.758 | 1.78 | 6 | 0 | [16 floats] |
| m_d411f369ab593053__f002 | 1 | 0.25 | 5.80078 | 0.812288 | 0.787 | 1.773 | 7 | 0 | [16 floats] |
| m_d411f369ab593053__f003 | 2 | 0.5 | 5.98828 | 0.811953 | 0.737 | 2.459 | 8 | 0 | [16 floats] |
Westfield North County · westfield_north_county
Escondido, CA · 33.07106, -117.06581 · walked 2026-09-20 by contributor_5 on iPhone16,1 (OS 26.6.2), landscape 1.0x · 333 clips, 17 min of video over 116 min
Track: time-gradient trail start → end on satellite imagery |  Coverage: trajectory coloured by frame-to-frame overlap, per-frame overlap / parallax strip |
 Examples: frames along the walk with their metric depth | |
15:54 into the walk · 5.0 s · m_0565d6a02907903f.mp4 | 39:42 into the walk · 5.0 s · m_9be9ab051eee5207.mp4 | 58:07 into the walk · 5.0 s · m_ed389cd0287967ba.mp4 | 90:28 into the walk · 5.0 s · m_3a4ef4e3e525e09c.mp4 |
De-identified clips (faces and plates blurred, re-encoded, no audio) (960x540), 206 in this walk under westfield_north_county/videos/.
| Capture | Coverage | ||
| clips (video / photo) | 333 (206 / 127) | video frames posed @ fps / shard samples | 4,110 @ 4 fps / 4,237 (incl. 127 photos) |
| video | 17.1 min | path length | 1,058.3 m |
| median clip | 5.00 s | footprint (p95 / max radius) | 157.3 m / 162.7 m |
| first → last upload (UTC) | 22:10:32 → 00:06:32 | vertical extent | 5.6 m |
| capture / uploaded size | 1920x1080 / 960x540 | GPS hull area | 21,495 m² |
| phone / OS / app | iPhone16,1 / 26.6.2 / 1.1.17 | revisits found / loops closed | 1 / 1 (RMS 0.24 m) |
| Depth and multi-view | GPS and georeferencing | ||
| median depth (p10 / p90) | 7.2 m (5.4 m / 9.7 m) | GPS quality | precise, median accuracy 19.7 m |
| valid depth per frame | 0.81 | fixes | 333 |
| overlap to next frame (median / p10) | 0.86 / 0.79 | altitude (median) | 109.6 m |
| parallax to next frame | 1.23° | reconstruction | pose graph with GPS edges, median residual 10.7 m, scale x1.05 |
| covisible frames | 16 | ||
| depth map size / FOV | 518x294 · 65° x 40° |
westfield_north_county/clips.parquet (333 rows):
| clip_id | created | media | duration_s | width | height | lat | lng | gps_accuracy_m | altitude_m | n_frames | media_path |
|---|---|---|---|---|---|---|---|---|---|---|---|
| m_08ad384633ec598b | 2026-09-20 22:10:32Z | video | 4.998 | 1920 | 1080 | 33.0715 | -117.065 | 31 | 107.9 | 20 | westfield_north_county/videos/m_08ad384633ec598b.mp4 |
| m_a1ed5b40d83d89f1 | 2026-09-20 22:11:08Z | video | 5 | 1920 | 1080 | 33.0713 | -117.066 | 20.7 | 108.8 | 20 | westfield_north_county/videos/m_a1ed5b40d83d89f1.mp4 |
| m_0f99b41411490e05 | 2026-09-20 22:11:42Z | video | 5 | 1920 | 1080 | 33.0714 | -117.066 | 23.7 | 110 | 20 | westfield_north_county/videos/m_0f99b41411490e05.mp4 |
westfield_north_county/frames.parquet (4,110 rows):
| frame_key | walk_frame_idx | t_rel_s | median_depth_m | valid_frac | overlap_next | parallax_deg | n_covisible | window_id | c2w |
|---|---|---|---|---|---|---|---|---|---|
| m_08ad384633ec598b__f001 | 0 | 0 | 7.33654 | 0.906449 | 0.955 | 0.087 | 8 | 0 | [16 floats] |
| m_08ad384633ec598b__f002 | 1 | 0.25 | 7.27929 | 0.906358 | 0.91 | 0.612 | 9 | 0 | [16 floats] |
| m_08ad384633ec598b__f003 | 2 | 0.5 | 7.08708 | 0.902602 | 0.904 | 0.659 | 10 | 0 | [16 floats] |
The Shops at Hudson Yards · hudson_yards
New York, NY · 40.75362, -74.00143 · walked 2026-09-20 by contributor_6 on iPhone18,2 (OS 26.6), portrait 1.0x · 369 clips, 25 min of video over 109 min
Track: time-gradient trail start → end on satellite imagery |  Coverage: trajectory coloured by frame-to-frame overlap, per-frame overlap / parallax strip |
 Examples: frames along the walk with their metric depth | |
20:22 into the walk · 5.0 s · m_3bfcc1161001b978.mp4 | 46:30 into the walk · 5.0 s · m_1463f797c176ffc9.mp4 | 70:23 into the walk · 5.0 s · m_f5628588a49ac466.mp4 | 94:07 into the walk · 5.0 s · m_8ac09681966309d9.mp4 |
De-identified clips (faces and plates blurred, re-encoded, no audio) (540x960), 301 in this walk under hudson_yards/videos/.
| Capture | Coverage | ||
| clips (video / photo) | 369 (301 / 68) | video frames posed @ fps / shard samples | 6,020 @ 4 fps / 6,088 (incl. 68 photos) |
| video | 25.0 min | path length | 2,134.1 m |
| median clip | 5.00 s | footprint (p95 / max radius) | 70.5 m / 80.8 m |
| first → last upload (UTC) | 21:11:24 → 22:59:58 | vertical extent | 14.3 m |
| capture / uploaded size | 1080x1920 / 540x960 | GPS hull area | 23,210 m² |
| phone / OS / app | iPhone18,2 / 26.6 / 1.1.17 | revisits found / loops closed | 5 / 4 (RMS 1.18 m) |
| Depth and multi-view | GPS and georeferencing | ||
| median depth (p10 / p90) | 8.0 m (6.6 m / 15.5 m) | GPS quality | precise, median accuracy 13.9 m |
| valid depth per frame | 0.89 | fixes | 369 |
| overlap to next frame (median / p10) | 0.87 / 0.67 | altitude (median) | 17.2 m |
| parallax to next frame | 1.64° | reconstruction | pose graph with GPS edges, median residual 26.2 m, scale x1.45 |
| covisible frames | 16 | ||
| depth map size / FOV | 294x518 · 33° x 55° |
hudson_yards/clips.parquet (369 rows):
| clip_id | created | media | duration_s | width | height | lat | lng | gps_accuracy_m | altitude_m | n_frames | media_path |
|---|---|---|---|---|---|---|---|---|---|---|---|
| m_e4924a6469762617 | 2026-09-20 21:11:24Z | photo | - | 2,000 | 1,500 | 40.755 | -74.0007 | 48.8 | -29.2 | 1 | hudson_yards/photos/m_e4924a6469762617.jpg |
| m_3c932f71661e4695 | 2026-09-20 21:11:55Z | video | 5 | 1,080 | 1,920 | 40.755 | -74.0006 | 44 | -25.3 | 20 | hudson_yards/videos/m_3c932f71661e4695.mp4 |
| m_47cd0a9a9d31934d | 2026-09-20 21:12:21Z | video | 4.998 | 1,080 | 1,920 | 40.755 | -74.0007 | 41.4 | -23.8 | 20 | hudson_yards/videos/m_47cd0a9a9d31934d.mp4 |
hudson_yards/frames.parquet (6,020 rows):
| frame_key | walk_frame_idx | t_rel_s | median_depth_m | valid_frac | overlap_next | parallax_deg | n_covisible | window_id | c2w |
|---|---|---|---|---|---|---|---|---|---|
| m_3c932f71661e4695__f001 | 0 | 0 | 4.00307 | 0.00265936 | 0.099 | 7.394 | 0 | 0 | [16 floats] |
| m_3c932f71661e4695__f002 | 1 | 0.25 | 4.08087 | 0.000275786 | 0.381 | 2.279 | 3 | 0 | [16 floats] |
| m_3c932f71661e4695__f003 | 2 | 0.5 | 3.7725 | 0.000735429 | 0 | 10.875 | 2 | 0 | [16 floats] |
Access and citation
This repository is a sample. For the full collection (more malls and cities, high-resolution media, IMU data), contact info@collectivememory.ai. License: CC BY 4.0.
@misc{cmusmalls2026, title={US Malls: sequential indoor mall walks with camera poses and metric depth},
author={Collective Memory}, year={2026}, howpublished={Hugging Face dataset collectivememory/us-malls-public}}
- Downloads last month
- 181