Hugging Face's logo

Datasets:
Dhi-Technologies
/
multicam-reasoning-memory-benchmark

Size:
100K<n<1M
Tags:
multi-camera
re-identification
tracking
memory
provenance
refusal-gate
License:
Dataset card Files
xet
 Community

Acknowledge the terms to access this dataset

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

Access is granted for non-commercial research and evaluation use only. By requesting access you agree not to redistribute this dataset or any derivative data, to cite Dhi Technologies in any publication or output that uses it, and to obtain a separate commercial license via dhi-tech.com before any commercial use. Access requests are reviewed manually by Dhi Technologies.

Log in or Sign Up to review the conditions and access this dataset content.

Multicam Reasoning Memory Benchmark

Product: multicam-reasoning-memory ("fleetmind"), "your cameras think together and remember": cross-camera identity linking with transit-time priors, weeks-scale bounded memory, and provenance-carrying queries.

Synthetic data, real-data validation in progress

This is a procedurally generated multi-site suite: 24 independent synthetic sites, each a chain of 3 to 8 cameras, where simulated people walk the camera chain with per-camera dwell plus transit jitter, emitting per-camera track events exactly like a real per-camera tracker would (a new track ID each time a person reappears on a different camera). Ground truth (which track IDs belong to the same physical person) ships alongside every site so linking precision and recall are measured, not assumed, for each site independently and in aggregate. No real multi-camera footage or real people are involved, and no public re-identification dataset (for example WILDTRACK or AI City) has been run through this product.

What's in this dataset

  • events.jsonl (165,902 rows, about 30.2 MB): track events from 24 synthetic sites, one row per event, each row tagged with a site_id so the sites can be filtered back apart. Schema, one JSON object per line:
    {"t": 28.84, "camera_id": "cam0", "track_id": "t27", "class": "person",
 "position_m": [9.57, 16.40], "identity_ref": "ref:person8", "site_id": "nc4_np20_rf0.2_seed0"}
    identity_ref is present only for the fraction of people carrying one, same as before, now varying per site (20 percent to 60 percent, see the grid below).
  • ground_truth.json (about 161.6 KB): a dict keyed by site_id. Each value carries that site's person_tracks (which camera:track_id keys belong to the same physical person), the true mean inter-camera transit times (transit_means_s), and generation metadata (meta, the camera count, people count, ref_fraction, and seed).
  • linking_report.json (new file, about 13 KB): one entry per site with n_events, precision, recall, linked_pairs, true_pairs, and the memory footprint (entities, episodes, links_active, relations, db_bytes), plus an aggregate block with mean precision and recall overall, the min and max recall observed, and breakdowns by camera count and by ref_fraction.
  • demo_seed1_report.json (about 480 bytes, unchanged): the original single-site, seed=1, 4-camera evidence file, kept as is for back compatibility with the example in this card's original "how to load it" walkthrough.

How to load it 

import json
from huggingface_hub import hf_hub_download

repo_id = "Dhi-Technologies/multicam-reasoning-memory-benchmark"
events_path = hf_hub_download(repo_id, "events.jsonl", repo_type="dataset")
gt_path = hf_hub_download(repo_id, "ground_truth.json", repo_type="dataset")
report_path = hf_hub_download(repo_id, "linking_report.json", repo_type="dataset")

events = [json.loads(line) for line in open(events_path)]
ground_truth = json.load(open(gt_path))
report = json.load(open(report_path))

print(len(events), "events across", len({e["site_id"] for e in events}), "sites")
print(events[0])
print(report["aggregate"])

Site grid and measured results

The suite spans 6 camera counts (3 through 8) and, for each camera count, 4 people or ref_fraction or seed variants, for 24 sites total. Every site was generated with fleetmind.synth.generate_site and scored by ingesting its events through CrossCameraLinker and MemoryStore, then comparing the resulting track_entity mapping to the site's own ground truth with score_linking, the same pipeline the repository's own test_site_linking_precision_over_recall test exercises.

site_id cameras people ref_fraction seed events precision recall
nc3_np20_rf0.2_seed0 3 20 0.2 0 1,442 1.0000 0.0476
nc3_np50_rf0.4_seed1 3 50 0.4 1 3,859 1.0000 0.3000
nc3_np80_rf0.5_seed2 3 80 0.5 2 6,028 1.0000 0.4699
nc3_np100_rf0.6_seed3 3 100 0.6 3 7,151 1.0000 0.6250
nc4_np20_rf0.2_seed0 4 20 0.2 0 1,715 1.0000 0.2500
nc4_np50_rf0.4_seed1 4 50 0.4 1 4,216 1.0000 0.3836
nc4_np80_rf0.5_seed2 4 80 0.5 2 7,047 1.0000 0.4167
nc4_np100_rf0.6_seed3 4 100 0.6 3 8,818 1.0000 0.5457
nc5_np20_rf0.2_seed0 5 20 0.2 0 2,098 1.0000 0.2389
nc5_np50_rf0.4_seed1 5 50 0.4 1 5,492 1.0000 0.3736
nc5_np80_rf0.5_seed2 5 80 0.5 2 8,436 1.0000 0.3416
nc5_np100_rf0.6_seed3 5 100 0.6 3 9,987 1.0000 0.6452
nc6_np20_rf0.2_seed0 6 20 0.2 0 2,541 1.0000 0.1183
nc6_np50_rf0.4_seed1 6 50 0.4 1 6,014 1.0000 0.4624
nc6_np80_rf0.5_seed2 6 80 0.5 2 10,407 1.0000 0.4581
nc6_np100_rf0.6_seed3 6 100 0.6 3 11,631 1.0000 0.5676
nc7_np20_rf0.2_seed0 7 20 0.2 0 2,684 1.0000 0.1562
nc7_np50_rf0.4_seed1 7 50 0.4 1 6,579 1.0000 0.4790
nc7_np80_rf0.5_seed2 7 80 0.5 2 9,785 1.0000 0.5171
nc7_np100_rf0.6_seed3 7 100 0.6 3 11,901 1.0000 0.5643
nc8_np20_rf0.2_seed0 8 20 0.2 0 2,925 1.0000 0.3938
nc8_np50_rf0.4_seed1 8 50 0.4 1 7,830 1.0000 0.3439
nc8_np80_rf0.5_seed2 8 80 0.5 2 11,724 1.0000 0.4903
nc8_np100_rf0.6_seed3 8 100 0.6 3 15,592 1.0000 0.6788

Aggregate across all 24 sites: mean precision 1.0000 (precision never dropped below 1.0000 on any site, the uniqueness guard held on every configuration tried), mean recall 0.4112, with recall ranging from 0.0476 (the smallest site, nc3_np20_rf0.2_seed0, 3 cameras and 20 people at 20 percent ref_fraction, where a low identity_ref rate combined with a short camera chain gives the linker very little to work with) to 0.6788 (the largest site, nc8_np100_rf0.6_seed3, 8 cameras and 100 people at 60 percent ref_fraction). Recall is not monotonic in camera count alone (compare nc6_np20_rf0.2_seed0 at 0.1183 with the neighboring nc5_np20_rf0.2_seed0 at 0.2389 and nc7_np20_rf0.2_seed0 at 0.1562), a real effect of seed variance at small people counts, not smoothed over here.

Grouped by camera count (mean recall, 4 sites each): 3 cameras 0.3606, 4 cameras 0.3990, 5 cameras 0.3998, 6 cameras 0.4016, 7 cameras 0.4292, 8 cameras 0.4767. More cameras give more chances for a person to accumulate confidently linkable transit pairs, so recall trends upward with camera count.

Grouped by ref_fraction (mean recall, 6 sites each): 0.2 ref_fraction 0.2008, 0.4 ref_fraction 0.3904, 0.5 ref_fraction 0.4490, 0.6 ref_fraction 0.6044. Recall tracks ref_fraction closely, as expected, since an identity_ref link is exact and does not depend on the transit-time prior at all.

This confirms, at wider scale than the original single seed, the same asymmetry the repository's regression test asserts: precision stays at or above 0.95 (here, exactly 1.0 on every one of the 24 sites tried) while recall is genuinely variable and depends on camera count, crowd size, and how many people carry an external identity_ref, not a fixed constant.

The original single seed, single site example (4 cameras, 20 people, 40 percent ref_fraction, seed=1, precision 1.0, recall 0.377, 61,440 byte memory footprint) is preserved unchanged in demo_seed1_report.json and is one specific point inside this wider grid (compare nc4_np50_rf0.4_seed1, the closest grid point, at recall 0.3836).

Regeneration provenance

Regenerated 2026-07-09 from DHI-Technologies-Inc/multicam-reasoning-memory at commit 2a156a2 (branch main). Grid: n_cameras in {3, 4, 5, 6, 7, 8}, and for each camera count, four (n_people, ref_fraction, seed) variants: (20, 0.2, 0), (50, 0.4, 1), (80, 0.5, 2), (100, 0.6, 3), for 24 sites total. All other generate_site parameters (dwell_s=30.0, transit_s=45.0, transit_jitter_s=6.0, event_period_s=1.0, start_spread_s=600.0) were left at their defaults, see Limitations below. Each site was produced with:

from fleetmind.synth import generate_site, score_linking
from fleetmind.linker import CrossCameraLinker
from fleetmind.memory import MemoryStore

truth = generate_site(n_cameras=nc, n_people=n_people, ref_fraction=ref_fraction, seed=seed)
store = MemoryStore(db_path)
linker = CrossCameraLinker(store)
track_entity = linker.ingest_all(truth.events)
scores = score_linking(truth, track_entity)

The full 24-site generate-plus-ingest-plus-score run completed in 16.6 seconds on the machine used to build this card. Total events summed to 165,902, comfortably inside the review budget (well under the 300,000-event and 150 MB thresholds that would have triggered thinning the grid), so the full requested grid ran with no thinning.

Schema notes

  • site_id (new) identifies which of the 24 generated sites an event or ground-truth entry belongs to. It encodes the site's camera count, people count, ref_fraction, and seed, for example nc8_np100_rf0.6_seed3 is 8 cameras, 100 people, 60 percent ref_fraction, seed 3.
  • position_m is carried through as provenance only; the linker itself does not use spatial position for scoring, only per-camera-pair transit time.
  • Timestamps are assumed clock-aligned (NTP) across cameras; there is no per-camera clock-offset estimation in the linking logic.
  • identity_ref is optional. When present, linking for that entity is exact (a plate, badge, or face-free cross-link from another product); when absent, the linker has to earn the link from timing alone, which is why overall recall is well below 1.0 across every site in the grid.

Method card, no trained weights

No trained model or weights ship with this product. Cross-camera linking is a transit-time prior plus a scored decision with a uniqueness guard (a refusal gate: if a second candidate is even plausible, the link is refused and a new entity is created instead). On the original synthetic site, margin alone gave 0.918 precision (a number attributed to the product's own source-code comment, not independently re-derived for this card); the uniqueness guard took it to 1.0 by trading recall, a deliberate, measured choice, since a false merge poisons weeks of memory while a missed link only costs a duplicate row. Across all 24 sites in the wider grid above, that same guard held precision at exactly 1.0 on every configuration tried. Read the deep-dive post.

Limitations

  • Synthetic only: no real multi-camera footage and no public re-identification dataset have been run through this product.
  • Recall (0.0476 to 0.6788 across the grid, mean 0.4112) is deliberately partial by design, not a ceiling the linker is trying and failing to clear: concurrent travelers without an identity_ref at the same camera-pair transit intentionally trip the uniqueness guard, trading a duplicate entity row for avoiding a wrong merge.
  • The benchmark now spans 24 sites across 6 camera counts (3 to 8) and 4 seeds (0 to 3), with people counts from 20 to 100 and ref_fraction from 0.2 to 0.6, so it is no longer a single seed or a single site topology. It still does not vary the per-camera dwell time, the inter-camera transit-time mean or jitter, the event emission rate, or the start-time spread, all 24 sites use the same dwell_s=30.0, transit_s=45.0, transit_jitter_s=6.0, event_period_s=1.0, and start_spread_s=600.0 defaults, only camera count, people count, ref_fraction, and seed were varied.
  • The linker does not use appearance, position, or any signal beyond transit time and the optional identity_ref; a real deployment with unreliable clocks or highly variable walking speed would need validation this synthetic benchmark does not provide.

License

This dataset is released under CC BY-NC 4.0 (non-commercial). Access is gated and requires manual approval: it is provided for non-commercial research and evaluation only, redistribution is not permitted, and any publication or output using it should cite Dhi Technologies. Commercial use requires a separate agreement, contact dhi-tech.com.

Try it

Source & research context

Downloads last month
9
Total file size:
31.8 MB

Space using Dhi-Technologies/multicam-reasoning-memory-benchmark 1

Collections including Dhi-Technologies/multicam-reasoning-memory-benchmark