cm-gaza

Verified in-the-wild phone footage with capture metadata, from a worldwide network of mappers · first regional release: Gaza · v0.1 preview · Collective Memory

The global corpus

Collective Memory runs an active, worldwide network of mappers: individuals who document places and events where they are, through an app that records the account, device time and GPS at the moment of capture and checks the media and its metadata on submission. As of 2026-09 the network had uploaded 18.6M clips from 174 countries, and it adds to them daily.

18.6M clips (3.0M video) from 172,792 contributors in 174 countries, through 2026-09. One dot per ~5 km cell of geotagged clips, sized and coloured by clip count (log scale); teal points are the last upload location of the 102,863 of 156,675 mappers who uploaded in the past 12 months. Per-cell counts in assets/global/candidates.csv. The Gaza corpus is 0.77% of the whole.

What distinguishes this data from web-scraped imagery is provenance at capture time and multi-view redundancy: the same places filmed by many people, from many phones, across months and years. That is the data that the following problems need and rarely get in-the-wild:

• Visual place recognition and localization across contributors, devices and dates, with verified pixel overlap as ground truth and GPS only as a coarse, sometimes absent, prior.
• Revisit-robust matching and change detection: the same street across visits during which it has changed.
• Multi-view 3D from uncalibrated phone footage: feed-forward and global structure-from-motion on mixed cameras, registered to the world.
• World-model and video-model training on real, verified, geolocated footage with camera trajectories recoverable from the data itself.

The global corpus is released in regional cuts that share one pipeline, one identifier space and one card structure, so that results transfer across them. This is the first cut; the next regional cut is in preparation. No single cut is representative of the global corpus; each is chosen because it is a hard, well-covered case for the problems above, and a model trained on one cut alone inherits that region's scene statistics (see Considerations).

This release: Gaza

Content warning. The rest of this card, and the data, document a war zone. Frames show destruction, injury, the dead and the wounded, grief and crowds under duress. Faces and licence plates are blurred; nothing else is edited. Review Considerations before use.

For the past two years (2024-06 → 2026-09) a group of mappers in Gaza has documented the Strip continuously: 143,138 clips — 125,133 photos and 18,005 videos (~5 s each, 1,422 min) from 93 contributors, still growing. Every clip is verified in-the-wild phone media: account, device time and GPS recorded at capture, integrity-checked on submission, and cross-matched against other footage of the same place. Together, these observations create a much richer record than any single camera, photographer or official feed could provide.

Verified observation of this kind — capture time, place and author recorded at the moment of capture, content confirmed by independent footage of the same place — makes it possible to build an immutable record of events as they unfold. It matters most in conflict zones, where coverage is limited or censored and the record is otherwise assembled after the fact from unverifiable sources. Gaza is such a case, and its conditions set the research problems this release carries:

• Positioning without reliable GPS. Phone GPS in the Gaza corpus is coarse (~100 m), intermittent and sometimes wrong, and 7,195 clips have none. Which clips see the same place, and where each camera was, must be computed from the images themselves; GPS is used only to choose where to look.
• Time differences with a changing landscape. The same street is filmed across months during which the surroundings are destroyed, cleared or rebuilt. Footage from different dates often shares only part of its geometry, so matching, registration and date-based hold-outs must cope with a scene that changes dramatically between visits.

This release (v0.1) is a preview of the Gaza corpus, exported 2026-08-24: 35,659 clips — 33,626 photos and 2,033 videos (163 min, 292k video frames at up to 30 fps) — from 47 contributors, 2024-07 → 2026-08, 25% of the Gaza corpus. It ships the raw, de-identified media in full with its capture metadata, a clip graph of pixel matches between the released clips, and one worked structure-from-motion example to show what the raw data supports. The frames table (43,357 rows: every photo, plus ~1 frame/s of every video) exists so the release can be browsed and streamed row by row; the videos themselves are under corpus/videos/. Faces and plates are blurred; nothing else is processed, selected or filtered.

We are releasing this cut now to invite research labs to work with us on the processing side of the data: placing footage that has no GPS from pixels alone, reconstructing sites from mixed phone cameras across dates during which the scene is destroyed or cleared, registering footage against a landscape that changes between visits, learning from verified in-the-wild video for world models. If you want to work on it, open a discussion on this repository or write to info@collectivememory.ai. We plan further releases from the corpus as review and de-identification are completed.

5 Gaza City video clips as shipped in corpus/videos/ (de-identified with face tracks, H.264, ≤ 1024 px, no audio; ~5 s each, press play).

Photos as shipped (de-identified, ≤ 1024 px): one from Khan Yunis, two from the unknown-location part. Click for full size.

What is released

The release on the Strip: every located clip of the Gaza corpus as grey density, the two released areas in blue, the SfM example as a star, and the layers of the release. Clips without a GPS fix have no position and are not drawn.

subset	clips	of which videos	rows in frames	what
sample	120	0	120	the SfM example's input clips (Al-Seqaly Street, Khan Yunis); first rows of the frames table
area	28,344	993	32,047	every located clip within 3.6 km of the example site (Khan Yunis, 26,985) and in two geohash-6 blocks of Gaza City (sv8e2h, sv8e2n, 1,359)
unknown_location	7,195	1,040	11,190	every clip of the Gaza corpus whose capture had no GPS fix; position unknown

Located parts (sample, area): 19.9% of the Gaza corpus, 38 contributors, 2024-07 → 2026-08. GPS is quantised to three decimals (~100 m) at capture; lat, lng, gh7 are those values, unmodified. Both areas are cut from the Gaza corpus by position only; nothing inside them was removed or selected. The Khan Yunis area is centred on the example site (31.3455 N, 34.3030 E).

Unknown-location part. When the phone has no fix the app records no coordinates. These 7,195 clips (1,040 videos) from 37 contributors ship with lat, lng, gh7, city null and gps_status = none. They belong to the Gaza corpus because their contributors otherwise upload almost exclusively from the Strip (inclusion_rule = account); they were captured under conditions in which GPS availability and reliability may be degraded, and absence of a fix does not establish the cause. Their content is the most violent in the release. They are included because geolocating media by means other than GPS is one of the research interests in the wider corpus.

Pixel placement. Pixel matching was run over the whole corpus, so unknown-location clips carry verified matches into located clips; the shipped graph keeps the pairs whose both ends are in this release. 2,376 of the 7,195 (33%) have at least one verified pixel match to a located clip of the released areas (placed_by_pixels); 6,207 belong to a cluster. For those, pixel_match_clip, pixel_match_inliers, pixel_match_lat, pixel_match_lng give the strongest located partner and its position — a hint, not a fix. Whether the rest can be placed, and how well, is one of the open questions this release is meant for. graph/clip_pairs.parquet has every verified pair between released clips.

Gaza corpus

The release is a 25% cut of the Gaza corpus: 143,138 clips from 93 contributors, 2024-06 → 2026-09, with a pixel-match graph of 2,548,868 clip pairs and 1,272 pixel clusters. 94 % of its clips carry a quantised fix; 5 % have none (released here in full); ~1 % were quantised to an offshore grid point and are not in this release. Documentation continues daily.

Clusters of the Gaza corpus, placed by GPS (alignment run v4). A cluster is a community of the pixel-match graph among clips whose fixes lie within 300 m (Louvain, resolution 20); clips without a GPS fix (7,195, 5.0% of the corpus) are attached to the cluster of their strongest pixel match and counted but not placed. Background: density of located clips with at least one verified pixel match (135,334 of 143,138 clips verified; 135,943 located). 1,272 clusters, 273 with >= 20 clips, of which 270 have >= 2 contributors and 241 have a parallax share >= 0.5 (median triangulation angle >= 2 deg on at least half of their verified pairs: multi-viewpoint, reconstructable in 3D). Per-cluster table for the released clips in assets/gaza/clusters.csv (the map covers the whole corpus); full-resolution map in assets/gaza/heatmap_full.png. Basemap and Gaza outline: OpenStreetMap. Blue: the released areas; star: the SfM example.

Usage

from datasets import load_dataset

frames = load_dataset("collectivememory/cm-gaza", "frames", split="train", streaming=True)   # 1 row per de-identified frame, JPEG embedded
row = next(iter(frames)); row["image"].size, row["clip_id"], row["region"], row["gps_status"]

clips    = load_dataset("collectivememory/cm-gaza", "clips",      split="train").to_pandas()   # 1 row per clip; video_path for videos
clusters = load_dataset("collectivememory/cm-gaza", "clusters",   split="train").to_pandas()   # 1 row per pixel cluster
pairs    = load_dataset("collectivememory/cm-gaza", "clip_pairs", split="train").to_pandas()   # verified clip pairs

One part only — corpus/frames/ is Hive-partitioned by subset:

unknown = load_dataset("collectivememory/cm-gaza", data_files="corpus/frames/subset=unknown_location/*.parquet", split="train")
area    = load_dataset("collectivememory/cm-gaza", data_files="corpus/frames/subset=area/*.parquet", split="train")
sample  = load_dataset("collectivememory/cm-gaza", data_files="corpus/frames/subset=sample/*.parquet", split="train")

Unknown-location clips with a pixel match into the located areas, and their partners:

placed = clips[clips.placed_by_pixels == True]
partners = clips.set_index("clip_id").loc[placed.pixel_match_clip, ["lat", "lng", "gh7", "city"]]

Full videos (de-identified, H.264, long side ≤ 1024 px, no audio) — one file per video clip, clips.video_path:

from huggingface_hub import hf_hub_download, snapshot_download
vids = clips[clips.video_path.notna()]
path = hf_hub_download("collectivememory/cm-gaza", vids.video_path.iloc[0], repo_type="dataset")
root = snapshot_download("collectivememory/cm-gaza", repo_type="dataset", allow_patterns=["corpus/videos/*"])   # all of them

The worked example (poses, depth, dense cloud):

root = snapshot_download("collectivememory/cm-gaza", repo_type="dataset", allow_patterns=["example/*"])
# example/sv86kd4_street/metadata.csv, images/, annotations/<clip>/{poses.npy,intrinsics.npy}, depths/<clip>.zip, site/{cameras.parquet,sparse/,*.ply,geo_transform.json}

Dataset structure

corpus/
├─ frames/subset=sample/part-0000.parquet          the example's 120 input frames (first rows of the viewer)
├─ frames/subset=area/part-XXXX.parquet            Khan Yunis then Gaza City, clusters by size
├─ frames/subset=unknown_location/part-XXXX.parquet
├─ videos/<clip_id>.mp4                            every video clip in full, de-identified
├─ clips.parquet
├─ blur_report.parquet · blur_stats.json · video_report.parquet
graph/
├─ clip_pairs.parquet · clusters.parquet · clusters.csv · cluster_members.parquet
example/sv86kd4_street/
├─ metadata.csv · images/ · annotations/ · depths/ · site/
assets/  gallery/ · area_map.png · gaza/ · global/ · sv86kd4_street/

table	rows	one row per	key columns
corpus/frames/subset=*/part-XXXX.parquet (frames)	43,357	photo, or sampled video frame	node_id, clip_id, frame, media, image (JPEG), width, height, contributor, captured, region, gps_status, lat, lng, gh7, scene, cluster, example, split, subset, n_faces_blurred, n_plates_blurred
corpus/clips.parquet (clips)	35,659	photo or video	clip_id, contributor, captured, media, n_frames, duration_s, region, gps_status, lat, lng, gh7, city, timestamp_suspect, inclusion_rule, scene, p_outdoor, cluster, example, split, placed_by_pixels, pixel_match_clip, pixel_match_inliers, pixel_match_lat, pixel_match_lng, subset, video_path, n_faces_blurred, n_plates_blurred, n_face_tracks, n_plate_tracks
graph/clip_pairs.parquet (clip_pairs)	294,312	verified clip pair	clip_a, clip_b, frame_pairs, max_inl, sum_inl, cross_creator, cross_date, same_cell, kind, tri_med, tri_max, h_ratio_min, gap_days, tier, score, cluster
graph/clusters.parquet / .csv (clusters)	591	pixel cluster	clips, clips_released, located, contributors, days, videos, pairs, parallax_share, cross_contributor_share, tri_med_deg, lat, lng, spread_p90_m
graph/cluster_members.parquet	31,084	clip in a cluster	clip_id, cluster
example/sv86kd4_street/metadata.{csv,parquet} (example)	120	clip of the example	id, media, image path, annotation path, depth path, resolution, contributor, captured, gps cell, sceneType, registered, n_points3d, split

• clip_id — salted hash of the platform's memory id (m_ + 16 hex); node_id = <clip_id>/fNNN. contributor — salted hash of the account (c_ + 12 hex). Both are consistent across all tables and across Collective Memory releases.
• captured — device time at capture (UTC). timestamp_suspect flags clips whose timestamp is inconsistent with the upload session.
• region — khan_yunis | gaza_city | unknown. gps_status — quantized (3-decimal fix), fix (rare unquantised fix), none (no fix).
• inclusion_rule — bbox (fix inside the Gaza Strip) or account (no fix; contributor uploads ≤ 20 located clips outside the Strip or ≥ 90 % inside).
• scene, p_outdoor — zero-shot CLIP labels, video clips only.
• Clusters vs sites. A cluster is a graph object: a Louvain community of the pixel-match graph among clips whose fixes lie within 300 m (resolution 20, edge weight = max inliers), with no-fix clips attached to the cluster of their strongest located match. A site is a physical place. The example is one site; a cluster is evidence that a site exists and where. cluster is null for clips without a verified match. Graph from alignment run v4.
• Media. Photos ship as one frame; videos ship as sampled frames in the frames table (~1 frame/s, up to 6 per clip) and in full under corpus/videos/ (video_path; H.264, no audio, original frame rate up to 30 fps). Everything is long side ≤ 1024 px and re-encoded after de-identification. Per-frame face/plate counts and boxes: frames.n_faces_blurred / n_plates_blurred, corpus/blur_report.parquet; per-video track counts, frame count, fps and size: corpus/video_report.parquet.

Worked example: example/sv86kd4_street/

120 photo clips from 7 contributors over 19 capture days (2025-11 → 2026-03) on Al-Seqaly Street, Khan Yunis Camp, registered into one frame: per-clip intrinsics and extrinsics (annotations/<clip>/{intrinsics,poses}.npy, site/cameras.parquet, COLMAP site/sparse/), pose-conditioned metric depth (depths/<clip>.zip, 000.npz with depth, conf, K, c2w), a sparse cloud (63k points) and two dense clouds (COLMAP MVS 0.5M, MapAnything 2.1M points), scale from MapAnything's metric cameras, and a hand-refined WGS-84 placement (site/geo_transform.json). Fixed hold-out splits by contributor and by capture date (split, site/splits.json) allow predictions from train clips to be evaluated against independently captured views. All 120 input frames are in corpus/frames/subset=sample/ and carry example = group_0001. The example is feed-forward global SfM on mixed phone cameras, not a calibrated capture: one focal length per camera, no distortion model, 3.2 px mean reprojection error at 700 px width, scale from a metric-depth prior (±10 %).

Highest-confidence pose-conditioned depths: input frame, metric depth, confidence.

Dense cloud (COLMAP MVS) of the example site.

Dense cloud rendered as point density over satellite imagery with OSM streets; placement hand-refined against imagery (a few metres in plan, heading to ~1°).

Dataset creation

Source data and verification

All media was captured through the Collective Memory app by its contributors (mappers): individuals who document places and events where they are, worldwide, in a peer-to-peer community for which Collective Memory Labs provides the infrastructure. The app records the account, device time and GPS at the moment of capture — the contributor, captured and lat/lng/gh7 fields here — and the platform applies integrity checks to submitted media and metadata. Within the global corpus, content is then cross-checked against other footage of the same pixels (the clip graph). Contributors agree to use of their footage for documentation and research under the dataset license and are compensated for their work by community members and Collective Memory Labs. No media was scraped from third-party platforms.

Processing

1. Media. Photos: one JPEG (quality 95). Videos: the full clip, re-encoded with H.264 at its original frame rate (≤ 30 fps), audio and container metadata removed; plus ~1 frame/s (up to 6 frames) sampled into the frames table. All media long side ≤ 1024 px. Blurring is applied before encoding, on every video frame and every still.
2. De-identification. Two face detectors are run on every shipped frame — YuNet at two scales and a YOLOv8 face model (WIDER FACE) — at four rotations for stills and two for video frames, boxes unioned, plus a licence-plate detector (YOLO11). Detections in videos are linked into tracks across frames (IoU, gaps ≤ 5 frames interpolated, tracks extended ±3 frames) so a face stays blurred through brief misses. Boxes are grown 15 % and Gaussian-blurred (sigma = 0.2 × box, elliptical for faces). In this release 212,000 faces in 34,866 frames and 1,531 plates were blurred in the frames table, and 221,255 face tracks and 2,241 plate tracks across 2,033 videos. Detections are gated by box size and confidence so that large low-confidence boxes (tents, rubble, ground) are not blurred. A manual review of 100 random stills and 10 random videos from this release found no unblurred frontal face and no blurred non-face region at review resolution; detection is still automatic, and heavily occluded, very small or turned-away faces can be missed. Report misses through a discussion on this repository and the clip will be re-processed or removed. Originals are not distributed.
3. Clip graph. MegaLoc place-recognition descriptors propose candidate frame pairs across the whole corpus (diversity-capped per contributor and burst, GPS-gated where a fix exists, no gating for no-fix frames); SuperPoint + LightGlue + MAGSAC verify them, keeping a pair only if enough inliers survive a fundamental-matrix fit, with two-view geometry (triangulation angle, homography ratio) stored per pair.
4. Clusters. Louvain communities on the verified graph among located clips within 300 m; no-fix clips attached to the cluster of their strongest located match.
5. Example geometry. GLUEMAP (feed-forward Pi3 stars + global rotation / similarity averaging + track refinement) registers the 120 clips into one COLMAP model; COLMAP MVS and pose-conditioned MapAnything densify it; MapAnything's metric cameras set the scale; placement on WGS-84 by hand against imagery.

All labels (clusters, pixel matches, scene classes, poses, depth, splits) are automatic. There are no human annotations.

Considerations for using the data

Content. This is documentation of a war. Faces and plates are blurred; scenes are not otherwise altered, cropped or filtered for severity. Do not use the data to identify individuals. Consider the exposure of annotators and reviewers.

Representativeness and feature bias. This cut is one region under one set of conditions: a dense urban strip during a war. Scene statistics (rubble, tents, damaged structures, crowds) are those of the place and time, not of the global corpus, and a model trained on this cut alone will inherit them. Use it for what it is, a hard, well-covered case for localization, matching and reconstruction under change, and draw training decisions against the corpus-level statistics in The global corpus; further regional cuts follow.

Personal and sensitive information. Media depicts real people and real places in a conflict zone. Shipped media is de-identified (faces and licence plates, automatic detection; see Processing for the residual-miss caveat) and may not be used to identify individuals. Contributor and clip identifiers are salted hashes. Takedown requests from contributors or from people depicted: open a discussion on this repository or write to info@collectivememory.ai.

Additional information

Licensing and access

CC BY-NC-SA 4.0. Access is a one-click agreement to the terms above (a Hugging Face login is required, which is what keeps the media from being scraped anonymously). Other terms: info@collectivememory.ai.

Maintainers

Collective Memory Labs — info@collectivememory.ai · takedowns, de-identification misses and partnership enquiries through the same address or a discussion on this repository.

Versions

• v0.1 (2026-09-17) — preview: Khan Yunis area, two Gaza City blocks, the full unknown-location set, full videos, SfM example sv86kd4_street.

Citation

@misc{cmgaza2026,
  title  = {cm-gaza: verified in-the-wild phone footage of Gaza with capture metadata},
  author = {Collective Memory Labs},
  year   = {2026},
  url    = {https://huggingface.co/datasets/collectivememory/cm-gaza},
  note   = {v0.1, built 2026-09-17}
}