Datasets:

Dhi-Technologies
/

multicam-reasoning-memory-benchmark

Size:

Tags:

License:

Dataset card Files Files and versions

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.

Multicam Reasoning Memory Benchmark

Product: multicam-reasoning-memory ("fleetmind"): cross-camera identity linking with transit-time priors, bounded memory and provenance-carrying queries.

Paper: Dhi Labs paper 09, No Link Beats a Wrong Link: Precision-First Cross-Camera Memory with Governed Identity Refusal (manuscript in preparation).

Results status. The headline results below are taken from the paper 09 abstract. The full set of results, tables and per-condition breakdowns will be posted here when the manuscript is final.

What this dataset is

Procedurally generated multi-camera sites. Simulated people walk a chain of cameras with dwell and transit jitter, and each camera emits track events in the form a per-camera tracker produces (a new track ID each time a person reappears on another camera). Ground truth, which track IDs belong to the same person, ships with every site. The files in this dataset contain no real-world imagery and no real tracker output.

What's in this dataset

events.jsonl (165,902 rows, about 30.2 MB): track events from 24 synthetic sites, each row tagged with site_id. 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 that carry an external identity reference.
ground_truth.json (about 161.6 KB): keyed by site_id. Each value holds the site's person_tracks, the true mean inter-camera transit times (transit_means_s), and generation metadata (meta).
linking_report.json (about 13 KB): one entry per site, plus an aggregate block.
demo_seed1_report.json (about 480 bytes): the single-site, seed 1, four-camera evidence file.

The 24 sites span 3 to 8 cameras, 20 to 100 people, reference fractions from 0.2 to 0.6, and seeds 0 to 3. The per-site scores in linking_report.json are not the source of the paper 09 results below.

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")

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

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

Results (paper 09 abstract)

Seeded synthetic site. The precision-first linker reaches precision 1.0 at recall 0.377. A matched ablation attributes the recovered links to the identity-reference path: ten-seed recall is 0.407 with prior learning on or off, and 0.0 without identity references.

City-scale benchmark, identity-disjoint validation over 1,004 transit directions. Each direction is checked against a 0.95 Wilson bound on test precision. In-sample, the spread of a direction's learned transit prior and its count of plausible candidates rank directions by whether they reach test precision 0.95 (AUROC 0.69 for each; corridor-cluster 95 percent intervals 0.56 to 0.81 and 0.58 to 0.81). The paper maps the identifiability boundary of the transit prior.

Framing. The paper is a diagnostic study. Its measured contributions are the identifiability boundary of the transit prior and the refusal behaviour of the memory.

Method summary

A candidate link must pass four gates: governed identity evidence (the production default refuses bare identity-reference strings); a transit-time likelihood learned only from reference-confirmed links; a margin over the runner-up; and a uniqueness guard that refuses the link when a second candidate is plausible.

Limitations

The files in this dataset are synthetic only.
The paper 09 results depend on the protocols and gates stated in the manuscript.

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.