Card: labeled vs unlabeled video, replay results on four testcases

README.md

@@ -46,6 +46,7 @@ Planned, not yet released:
 | 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.
 
@@ -78,6 +79,28 @@ calibration/
 
 The calibration JSON (scenes, sensor distortion models, stereo extrinsics, ArUco cube definitions) is also bundled in the [multiview repository](https://github.com/prcvlabs/multiview) 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.
+
+```bash
+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:
@@ -107,6 +130,7 @@ To use the calibration assets from this dataset instead of the copy in the repos
 
 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.
@@ -130,7 +154,7 @@ The multiview pipeline (HyperPose pose detector, `--no-stereo`, `pose-svm-lite_v
 
 **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.
 
-**In development: replaying original poses.** multiview can now load 2D poses saved from a run on the original video instead of detecting them ([PR #11](https://github.com/prcvlabs/multiview/pull/11)). On 007, the blurred video with poses replayed from the original scores exactly what the original does (F1 0.614, MOTA 0.579, 0 ID switches). Publishing these pose files alongside the videos is planned.
+**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](https://github.com/prcvlabs/multiview/pull/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.