Test results · 4 November 2026
All 17 MultiFusion AI agents ship in Server 0.7.2. We installed the released server from the Download Center, set it up like a customer, and ran every agent on real footage. Each film below draws exactly what the server stored.
Test setup: the released MultiFusion Server 0.7.0 and Analytics Worker 0.2.0 (then upgraded in place to 0.7.1 / 0.2.1 and 0.7.2 / 0.2.2), downloaded from the Download Center and installed on a clean Ubuntu 24.04 machine with 4 processor cores and no graphics card. Each clip played as a camera; the server recorded it; each agent then analysed the server's recording at its full frame rate. Counts are rows the test server stored. Accuracy is AP50 (average precision at 50% overlap) unless stated.
Before testing the agents we installed the released server exactly as a customer does, then exercised the whole system. Every row below happened on that machine.
| Step | What happened on the released 0.7.0 |
|---|---|
| Download | Server 0.7.0 (Linux x86-64) and Analytics Worker 0.2.0 from the Download Center; both packages match their published SHA-256 |
| Install | The package's own installer on a clean Ubuntu 24.04: database, media engine and 36 signed AI models installed with nothing downloaded; the server reports 0.7.0 healthy |
| Set up | The setup wizard's nine steps (prepare, network, services, storage, licence, security, administrator, start, validate) all completed |
| Licence | The 30-day trial, offline: 16 cameras, 10 users, every feature |
| Analytics Worker | Paired with a one-time code; online over a secure link; every model installed from the server |
| Cameras and recording | Test clips played as RTSP cameras and recorded continuously in 10-minute segments |
| AI agents | Every agent run on the server's own recordings at full frame rate; the films below draw what the server stored |
| Watch list | A watch list created, one photo enrolled and matched on video |
| Search | Footage searched by words (Recall, on 0.7.1); people found again across two camera views (Echo); plates searchable by number |
| Video synopsis | A camera's footage turned into a synopsis video with 164 objects, each at its real time |
| Signed evidence | A 16-second clip exported as a signed evidence package and checked with OpenSSL: signature verified, file hash matches the manifest; a changed copy fails |
| Audit | The hash-chained audit log verified intact (73 entries) |
| Upgrade | The 0.7.1 installer run over the installed 0.7.0: the server reports 0.7.1 healthy with every camera, recording, detection, watch list and the audit trail kept (189 entries, verified); the worker upgraded to 0.2.1 and installed the new model versions by itself |
The evidence package from this test, checked on the command line with OpenSSL alone, as a court or an auditor would:
$ openssl x509 -in ca.crt -noout -fingerprint -sha256 sha256 Fingerprint=42:23:5C:C9:B1:86:57:29:DC:4D:A3:49:94:B3:90:90:01:5F:08:CA:1B:69:A4:18:92:76:15:CD:BC:BB:D5:FC $ openssl verify -CAfile ca.crt signer.crt signer.crt: OK $ openssl dgst -sha256 -verify signer.pub -signature manifest.sig.der manifest.json Verified OK $ sha256sum 7e800d84-2335-48f2-b15a-a6d321b9bcca.mp4 4b5373bc02178bb395d41aed09bf8242f71e80a899b4fb17e7c56ff9eef3f6f6 7e800d84-2335-48f2-b15a-a6d321b9bcca.mp4 $ grep -A1 '"name": "7e800d84' manifest.json | tail -1 "sha256": "4b5373bc02178bb395d41aed09bf8242f71e80a899b4fb17e7c56ff9eef3f6f6", $ openssl dgst -sha256 -verify signer.pub -signature manifest.sig.der /tmp/m.json # manifest.json with one character changed Verification failure
People and vehicles, 80 object kinds
Sharper detection, Lite, Standard and Pro
Find and read number plates and text
Search footage by picture or by words
Find every face; match against watch lists
Follows each person and vehicle: counts, dwell, queues, wrong way, heatmaps
The same person across cameras
Body pose: falls and fights
Safety gear: helmets and gloves
Fire and smoke, confirmed over seconds
Riders without helmets, three on one bike, drivers without seat belts; each rider's head read close up
Clothing, bags, vehicle type and colour, for search
Counts dense crowds
Gunshots and breaking glass on the camera's microphone
A camera's hours shown as minutes
Analytics Worker 0.2.0's built-in benchmark (mfinfer bench), run on 2 and 3 November 2026, with the MultiFusion Sentinel detector at 640 × 640. Camera capacity counts the detector and video decoding only: plate reading and search cost more per camera. Every raw result file is linked.
| Machine | Runtime | Measured | Raw result |
|---|---|---|---|
| Laptop: Intel Core i7-13700HX with Intel UHD Graphics, Windows 11 | OpenVINO pack on the Intel UHD Graphics | Detector 45 ms, 17.8 fps (18.1 fps at FP16); 1 camera at 10 fps per detector (3 channels of five-type rules, or 8 crowd channels) | JSON |
| Same laptop: NVIDIA GeForce RTX 5050 Laptop GPU (8 GB), Ubuntu 24.04 on WSL2 | NVIDIA CUDA 12 pack, CUDA | Detector 7.7 ms, 76.5 fps | JSON |
| Same GPU, Ubuntu 24.04 on WSL2 | NVIDIA CUDA 12 pack, TensorRT, with video decoding on the GPU | Detector 151.5 fps (5.0 ms per frame at FP16); 15 cameras at 10 fps (30 channels of five-type rules, or 75 crowd channels) | JSON |
| Same laptop: Intel Core i7-13700HX processor, Ubuntu 24.04 on WSL2 | Analytics Worker 0.2.0 Linux runtime on the processor | Detector 54.7 ms, 17.8 fps; 1 camera at 10 fps per detector | JSON |
Test footage: scenes from MultiFusion's own film; clips from Pexels and Mixkit, whose free licences allow commercial use; real gunshot and glass recordings for Sonar; and held-out frames from public datasets where named. Accuracy figures are on each model's public held-out test set.
Want to see an agent on your own footage? Book a demo and bring a recording.
