prcvlabs/multiview-retail-tracking

multiview retail tracking

Test scenarios and calibration data for multiview, the open-source C++ library for multi-camera 3D people tracking written by Aaron Michaux at Perceive. The data comes from Perceive's ceiling-mounted stereo cameras in retail stores and offices. The project history is at prcvlabs.org.

Why this dataset

A real-world, ceiling-mounted stereo multi-camera tracking benchmark from deployed retail systems, released privacy-preserving, together with the open-source system that produced it. It is meant as a real-store test set, for example for trackers trained on synthetic data.

What sets it apart:

• Calibrated overhead stereo. Every camera is a ceiling-mounted stereo pair with fisheye calibration, so depth is available at every view. Most multi-camera tracking datasets use monocular cameras mounted at an angle.
• Real deployments. Recorded by a production system in operating retail stores and offices across several sites (2018 to 2020), with real lighting, clutter, crowding and camera re-mounts. Not synthetic or staged.
• Code and data from the same system. multiview, the tracker that ran in these spaces, is open source, and every scenario ships with its calibration and ground truth so baselines are reproducible.
• Privacy-preserving release. Faces and signage are blurred, and the effect on tracking is measured (below).

Planned, not yet released:

• 2D pose side files computed on the unblurred originals, so pose-based tracking on the blurred videos matches the originals (see below).
• Per-frame 2D keypoints from current pose models and stereo-triangulated 3D skeletons.
• Exports to standard multi-camera tracking formats, standard metrics (HOTA, IDF1), and a larger set of multi-camera scenes with new ground truth.

At a glance

Scenarios	82 directories under testcases/: 81 runnable scenarios with a manifest.json (013_site06_occlude holds only expected-output.text)
Sites	7 anonymized sites (site01 to site07): retail stores and showrooms, a public lobby, and an office
Videos	527 MP4 files, 126,719 frames, 8,442 s (2 h 21 min) of synchronized footage
Ground truth	45 scenarios have ground-truth.json (annotated floor tracks); 45 have annotation-data.json
Size	2.36 GB total: 2.15 GB of scenarios and 0.20 GB of calibration assets
Unlabeled video	84 segments under unlabeled/, 6.7 camera-hours, no ground truth (see below)

Every video frame is a side-by-side stereo pair: 1792x672 pixels, with the left and right sensors at 896x672 each. 523 videos are 15 fps and 4 videos (001_site01_short, 999_site01_example) are 25 fps. Scenario lengths range from 2.4 s to 117 s, and most site04 scenarios are 6 to 20 s across 9 cameras.

Site	Scenarios	Videos	Cameras per scenario	Setting
site01	4	8	2	retail showroom
site02	2	4	2	retail store
site03	2	12	6	retail store
site04	48	432	9	showroom (standing, sitting, lying)
site05	1	1	1	public lobby (crowds, steep angle)
site06	6	40	8	showroom
site07	19	30	1 to 2	office (lying, sitting, ladder, lobby)

Layout

testcases/<NNN>_<site>_<activity>/
    manifest.json          # read by `multiview-cli pipeline -m`; lists cameras, video files, fps, scene key
    C<camera>[_v<k>].mp4   # one side-by-side stereo video per camera (face-blurred)
    ground-truth.json      # annotated tracks (where available); also the input to --pipeline-testcase
    annotation-data.json   # raw annotation data from the multiview GUI (where available)
    expected-output.text   # pass criteria, e.g. "P>.95, R>.95, F1>.9"
    hash.md5               # md5 of ground-truth.json, used by the fetch script
calibration/
    scenes/  sensors/  binocular/  aruco-cubes/  aruco-results/  plane-data/
    camera-positioning-annotations/   # small JSON calibration (also bundled in the multiview repo)
    classifiers/*.ml.gz               # pose/tpfp random-forest classifiers (149 MB)
    cad-models/*.stl                  # store CAD models (38 MB)

The calibration JSON (scenes, sensor distortion models, stereo extrinsics, ArUco cube definitions) is also bundled in the multiview repository under data/calibration/, which is multiview's default asset store. It is duplicated here so the dataset is self-contained.

Labeled and unlabeled video

The dataset has two parts:

• testcases/: labeled. These are short scenarios with calibration. 45 of them have annotated floor tracks (ground-truth.json), so they can be scored.
• unlabeled/: no ground truth. This is longer real-store footage in the same format: side-by-side stereo, faces and signage blurred the same way. Use it for pretraining, self-supervised learning, detection and pose evaluation, and tracking without scores. Long idle stretches were cut. A segment keeps every stretch where any camera sees a person, plus 5 s on each side, and gaps under 20 s are kept. All cameras of a scene are cut to the same frame ranges, so they stay synchronized.

Site	Segments	Camera-hours	Cameras	Calibrated	Notes
site04	21	1.93	9-camera scene + single-camera clips	9-camera scene: yes (site04_scene05)	The 9-camera scene (site04_u01) runs in multiview with its manifest.json
site02	2	0.33	2	yes (site02_scene01_v1)	runs in multiview with its manifest.json
site05	56	4.25	2 monocular cameras (camA, camB)	no	public lobby; cut from a debug render that showed one eye of two cameras, so these are not stereo
site06	3	0.11	1	not included	single-camera clips
site08	2	0.10	1 (fisheye stereo)	not included	office, 2018

Each unlabeled/<segment>/ holds the videos, a meta.json (site, fps, frame count, cameras, whether it is calibrated) and, for calibrated scenes, a multiview manifest.json. unlabeled/index.json lists every segment.

hf download prcvlabs/multiview-retail-tracking --repo-type dataset \
    --include "unlabeled/site04_u01_s01/*" --local-dir ~/multiview-data

Using it with multiview

No login is needed. From a multiview checkout:

# Download one or more scenarios (by number) into ~/multiview-data/testcases/
# (set MULTIVIEW_TESTCASE_DATA to use another directory)
build/fetch-testcases.sh 007 011

# Or download directly with the Hugging Face CLI
hf download prcvlabs/multiview-retail-tracking --repo-type dataset \
    --include "testcases/007_*/*" --local-dir ~/multiview-data

# Run a scenario and score it against its ground truth
bin/testcase.sh 007_site03_sparse_tracks

# ...which amounts to:
multiview-cli pipeline -m ~/multiview-data/testcases/007_site03_sparse_tracks/manifest.json \
    -d /tmp/out --no-stereo --pose-classifier pose-svm-lite_v0 \
    --tracks-file ~/multiview-data/testcases/007_site03_sparse_tracks/ground-truth.json \
    --pipeline-testcase

To use the calibration assets from this dataset instead of the copy in the repository, download calibration/** and point multiview at the download directory with MULTIVIEW_ASSET_STORE=MULTIVIEW_ASSET_DIR MULTIVIEW_ASSET_DIR=<dir>. Note that the pipeline writes cache files (cached-data/) into the asset directory.

Privacy and anonymization

The data is freely downloadable because faces are heavily blurred and locations and dates are removed. Please use it only for non-commercial research and education, and do not attempt to identify any person or location shown.

• The unlabeled video was treated the same way. Detection used the same two models at 1280 px on every second frame, with the same head ellipses, ±3-frame union and blur strength (pixelate 10, Gaussian sigma 6), and was encoded as x264 CRF 16. Signage boxes were carried over from the published videos of the same camera by feature matching, or drawn by hand for cameras not published before. QA used the same two independent detectors as below, on the busiest frames plus a random frame of each of 150 output videos. They found 671 faces on the original frames and 58 on the blurred frames. Manual review of the 58 found no identifiable faces: they were blurred head blobs, patterned fabric, lights, a mirror and a fire-alarm sign, plus the tops of heads seen from directly above.
• Faces are blurred in every video. Heads were located in each stereo half separately with two detectors. The first is a YOLO11x-pose model (Ultralytics, 1280 px input, low confidence threshold), whose nose, eye, ear, and shoulder keypoints give the head position and size. The second is a YOLOv11l face detector trained on WIDER FACE, which catches small and background faces. Detections are linked into tracks, and short gaps (up to 12 frames) are interpolated. Each head is replaced with a strongly pixelated and Gaussian-blurred ellipse (about 1.1 times the head radius plus a margin), unioned over ±3 frames so moving heads stay covered. The body outline and head silhouette stay visible. multiview's own blur_faces command was not used, because it runs pose on the full double-width frame and does not write its output video.
• Signage is blurred. Store names, logos, slogans, logo floor mats, a poster face, and on-screen program listings were blurred with fixed per-camera boxes. The boxes were re-aligned to each video by feature matching, because cameras were remounted between some sessions.
• Video format is preserved. Videos were re-encoded as H.264 (CRF 16) in MP4 at the original resolution, frame rate, frame count, pixel format, and per-frame timestamps, so the pipeline reads them exactly like the originals. Container metadata was stripped.
• QA. Two face detectors that were not used for blurring (OpenCV YuNet and a YOLOv8m face model) were run on 132 frames from 22 videos covering all 7 sites, weighted toward the most crowded moments. They found 723 faces on the original frames and 75 on the blurred frames. Manual review of the 75 found no identifiable faces: about half were detections on blurred head blobs, and the rest were false positives on shelves, lamps, and furniture. Very small figures far behind shop windows (about 10 px tall) are not always blurred.
• Location names and dates are removed. Each store or office is identified only by a neutral site ID (site01 to site07). Scene keys, scenario names, file paths, and timestamps were rewritten consistently across all files, so scenes still load.

Effect of blurring on tracking

The multiview pipeline (HyperPose pose detector, --no-stereo, pose-svm-lite_v0 classifier, identical params.json) was run on the original and the blurred videos and scored against ground-truth.json with --pipeline-testcase. MOTA is 1 − (FN + FP + ID switches) / ground-truth points. A repeated run on the original videos of 011 gave identical numbers, so these differences are not run-to-run noise.

Scenario	Video	Precision	Recall	F1	MOTA	ID switches	GT / detected tracks
006_site03_tracks	original	0.935	0.852	0.805	0.793	0	13 / 4
	blurred	0.994	0.516	0.515	0.512	1	13 / 4
007_site03_sparse_tracks	original	0.880	0.670	0.614	0.579	0	2 / 5
	blurred	0.871	0.149	0.145	0.127	0	2 / 2
011_site06_standing	original	1.000	0.250	0.250	0.250	0	4 / 1
	blurred	0.570	0.076	0.071	0.018	0	4 / 1
080_site07_front_lobby_front_1	original	0.349	0.228	0.160	−0.197	0	3 / 6
	blurred	0.342	0.194	0.142	−0.179	0	3 / 6

Blurring substantially lowers recall in the multi-camera store scenes. HyperPose, a bottom-up OpenPose-style model, relies on head appearance to find people from overhead views, and with faces removed it misses many detections. Precision is largely unchanged. In a test on 007, a lighter Gaussian-only blur recovered recall to 0.300, and tighter masks without temporal dilation did not help (0.134). This confirms that the blur itself, not the mask size, causes the loss. Use these videos for development and relative comparisons. Results are not directly comparable to numbers obtained on the unblurred originals.

Replaying original poses. multiview can load 2D poses saved from a run on the original video instead of detecting them (pose_skeleton_init.op_params.replay_dir, PR #11). With poses replayed from the originals, the blurred videos score exactly what the originals do: F1 0.805 on 006, 0.614 on 007, 0.250 on 011 and 0.160 on 080, with no change in ID switches. Publishing these pose files alongside the videos is planned.

If you believe any frame still identifies a person or location, contact the maintainers and it will be removed.

License

The data is released under CC BY-NC 4.0. The multiview code is Apache 2.0.

Credits

• multiview was designed and written by Aaron Michaux at Perceive.
• Data collection and the dataset release are by PRCV Labs (Everett Berry).
• Perceive's research was funded by the U.S. National Science Foundation under SBIR Phase I Award 1622082 and Phase II Award 1738492.