Quality report (review sample)
Scope: the four sessions in this sample (days 1, 8, 13 and 15). Sections 1 to 5 restate numbers from the paper and from the dataset's existing quality-assurance runs over the whole window. Section 6 was computed fresh from the files of this sample and is marked as such.
1. Sessions
| day | weekday | type | hours | node-hours | live tracks | occupancy level |
|---|---|---|---|---|---|---|
| 1 | Tue | term weekday | 19.0 | 133.0 | 3,200 (from day 2 02:09 UTC only) | -- (partial) |
| 8 | Tue | low-occupancy weekday | 19.0 | 133.0 | 9,454 | moderate |
| 13 | Sun | weekend | 15.0 | 105.0 | 7,000 | moderate |
| 15 | Tue | term weekday | 19.0 | 133.0 | 25,763 | busy |
Occupancy level is set by the 90th percentile of live tracks present in the session: quiet up to 20, busy above 85, moderate in between. Each session ran from opening to 07:00 UTC the next day. All times in this report are UTC (shifted); a session's day number is the day it started.
2. Recording completeness (recorder manifest and archive verifier)
| day | nodes with frames | share of the span with all 7 nodes | single-node outages | restarts | verifier findings |
|---|---|---|---|---|---|
| 1 | 7 / 7 | 0.999997 | 0 | 0 | none (exit code 0) |
| 8 | 7 / 7 | 0.999998 | 0 | 0 | none (exit code 0) |
| 13 | 7 / 7 | 0.999995 | 0 | 0 | none (exit code 0) |
| 15 | 7 / 7 | 0.999997 | 0 | 0 | none (exit code 0) |
Intervals per node are in metadata/recorder/<session>/verification_findings.json (frames less than 1 s apart form
one interval).
3. Live tracklets (whole-window tracklet QA)
| day | verdict | reason |
|---|---|---|
| 1 | FAIL | the live log starts at day 2 02:09:46 UTC (25 % of the session span), after the benchmark walk; see section 5 |
| 8 | WARN | nearest-node share of sightings under N5 is 0.4 % against a median of 4.0 %; an occupancy effect of the low-occupancy period (the east seating is nearly empty), not node loss: every node recorded at full rate |
| 13 | WARN | nearest-node share under N7 is 4.6 % against a median of 9.7 %; weekend occupancy, as above |
| 15 | PASS | - |
Across all 21 live sessions, 7 tracks in about 306,000 show a sustained jump faster than 4 m/s, all while coasting; one of them is on day 15 at 15:40:36 UTC. Frame stamps are near-monotonic, and no live log has a gap longer than 5 minutes inside a session.
Known properties of the live tracklets:
- The live tracker restarts daily at 11:05 UTC; there are no frames from about 10:50 to 11:05 UTC, and track IDs restart at every restart. Key tracks by (tracker run, id).
- The live tracker ran at 2.5 Hz after its restart and slowed to 1.3 to 1.8 Hz by evening on the busiest days. A lost moving track coasts about 8 s at 2.5 Hz and about 15 s at 1.3 Hz.
- Live tracks last a median of about 35 s, because a person who stays an hour is often carried by several tracks.
4. Experimental flicker flags
The filter flags tracks born at a few fixed spots that coast for most of their life and barely move. It writes flags and deletes nothing; the raw tracklets are the default. Whole-session counts from the run over all 21 sessions (the filter counts each session's track set with its own boundary convention, so totals can differ by one from section 1):
| day | tracks | flagged |
|---|---|---|
| 1 | 3,201 | 436 |
| 8 | 9,454 | 2,561 |
| 13 | 7,001 | 1,538 |
| 15 | 25,764 | 2,323 |
Validation against manual headcounts: in the busy benchmark session the median number of tracks present moves from 88 to 87 with the flags applied, against a headcount of about 85 (toward it); in the quiet session from 9 to 8, against about 9 (away from it). The moderate session has no live log and no headcount. The filter is experimental and not accepted as a correction.
5. Benchmark sessions
| level | session day, time (UTC) | tracks present | walked | live rate |
|---|---|---|---|---|
| quiet | Sun, day 13, 16:02-16:41 | about 9 | 940 m | 2.5 Hz |
| moderate | Tue, day 1 (day 2 00:39-01:26) | about 70 | 932 m | 1.3 Hz (replayed) |
| busy | Tue, day 15, 17:30-17:55 | median 88 | 940 m | 2.5 Hz |
Reference results (walker-level; IDSW = ID switches per 100 m; Seat = seated stops after which the walker kept the ID):
| level | tracker | MOTA* | IDF1 | IDSW | Seat |
|---|---|---|---|---|---|
| quiet | live (2.5 Hz) | 0.97 | 0.93 | 0.32 | 6/6 |
| quiet | offline (10 Hz) | 0.99 | 0.97 | 0.21 | 5/6 |
| moderate | live (1.3 Hz, replayed) | 0.93 | 0.75 | 1.07 | 2/6 |
| moderate | offline (10 Hz) | 0.99 | 0.95 | 0.43 | 4/6 |
| busy | live (2.5 Hz) | 0.88 | 0.75 | 1.38 | 2/6 |
| busy | offline (10 Hz) | 0.94 | 0.85 | 0.85 | 1/6 |
A manual headcount during the busy session found about 85 people in the room. In the busy session two walkers walked
the four routes concurrently. The benchmark windows are tagged in sessions.csv and should be excluded from analyses
of everyday behavior.
The moderate session predates the live tracklet logger, so its live tracklets are the live tracker replayed on the recorded clouds at the live tick cadence; route B, also walked that evening, is not a benchmark route.
6. This sample (computed fresh from the package files)
Computed by the sample build from the files of this package (not numbers from the paper). Times are UTC.
Tracklet files
| file | tracker frames | track rows | distinct tracks | median rate | first - last (UTC) |
|---|---|---|---|---|---|
track_log/tracks_day8_000000.jsonl.gz |
213,688 | 2,781,569 | 10,617 (tracker run, id) | 2.50 Hz | 03:59:59 - 03:59:59 (next day) |
track_log/tracks_day15_000000.jsonl.gz |
165,373 | 8,312,597 | 26,452 (tracker run, id) | 2.08 Hz | 03:59:59 - 03:59:58 (next day) |
benchmark/busy/tracks_final_5hz.csv.gz |
7,494 | 692,019 | 1,572 ids | 5.00 Hz | 17:30:00 - 17:54:59 |
benchmark/busy/tracks_live_day15_133000.jsonl.gz |
3,901 | 342,792 | 1,040 (tracker run, id) | 2.50 Hz | 17:30:00 - 17:55:59 |
benchmark/quiet/tracks_final_5hz.csv.gz |
13,440 | 138,404 | 418 ids | 5.00 Hz | 16:00:10 - 16:44:59 |
benchmark/quiet/tracks_live_day13_120200.jsonl.gz |
6,000 | 53,333 | 254 (tracker run, id) | 2.50 Hz | 16:02:00 - 16:41:59 |
benchmark/moderate/tracks_final_5hz.csv.gz |
19,489 | 1,429,730 | 2,811 ids | 5.00 Hz | 00:30:00 - 01:34:59 |
benchmark/moderate/tracks_live_replayed_day1_203900.csv.gz |
3,738 | 279,667 | 1,193 ids | 1.30 Hz | 00:39:00 - 01:26:59 |
The offline files in offline/ are byte-identical copies of the benchmark offline files; each covers the whole offline
run, and window.json gives the benchmark window.
Point-cloud windows
| window | nodes | segments | frames (all nodes) | points | bytes |
|---|---|---|---|---|---|
clouds/geo_day8_080011 (19:00:00-20:00:00 UTC) |
7 | 49 | 254,899 | 990,441,953 | 11.89 GB |
benchmark/busy/clouds/geo_day15_080014 (17:30:00-17:56:00 UTC) |
7 | 28 | 110,610 | 474,966,718 | 5.70 GB |
benchmark/quiet/clouds/geo_day13_120009 (16:02:00-16:42:00 UTC) |
7 | 35 | 169,778 | 635,130,164 | 7.62 GB |
benchmark/moderate/clouds/geo_day1_080012 (00:39:00-01:27:00 UTC) |
7 | 42 | 203,979 | 856,542,452 | 10.28 GB |
Flicker flags in this sample: 5,775 flags of tracks born in the sample's live data (metadata/flicker/flags.tsv).