A Rust Object Detector on Your Home Computer
A Rust object detector on your home computer: Diana 0.11 does 34.6 percent less counted work per detect than 0.9.0. Host software only, untimed, and not wired to a camera.

Can my home computer run a Rust object detector for a camera?
The detector exists as host software: published ffai-diana 0.11.0 runs YOLO26 detection and ByteTrack tracking on a desktop-class CPU. It is not wired to a dldeploy camera stream, and its new 34.6 percent figure is counted instructions, not frames per second.
A Rust object detector on your home computer just got cheaper in counted work. Diana, the detection and tracking part of the FFAI suite, is published as ffai-diana 0.11.0 on crates.io. It carries the 0.10 campaign, which cut one detect on the smallest YOLO26 model from 1,075,156,773 instructions to 702,792,245, down 34.6 percent, against the 0.9.0 release. It is host software. It was not timed, it is not on any ESP32, and no dldeploy camera feeds it yet.
This page belongs to the series that starts at smart home access on your home computer. Each page names one job, what already runs on the LAN, and what is still unmeasured. The job here is noticing what is in a camera frame, a person, a car, a package, on the machine you own instead of in a vendor's cloud.
Every number below is repository-reported, from commit e119f1a9 and its ledger. We ran nothing for this page.
What an object detector on your home computer counted
Diana's campaign counted CPU instructions with callgrind, the Valgrind counting tool, under WSL. It pinned the work to one core and one thread, named rustc 1.95.0 and Valgrind 3.26.0, and did not name the CPU. Two workloads were measured.
The detect workload is the yolo26n model on one photo from the COCO dataset. One detect fell 34.6 percent in instructions. The tracker workload replays cached detections from the MOT17-02 sequence of the MOT17 benchmark, a public pedestrian-tracking video. One tracker replay fell from 21,336,089 to 16,148,683 instructions, down 24.3 percent.
For an object detector on your home computer, one photo and one cached tracking clip are a yardstick. They are not your porch at night, in rain, with a moth on the lens. The campaign does not claim to describe those.
The work came from not moving data twice: copying convolution rows with one memcpy, writing results straight into the buffers they are headed for, fusing small steps together, and flattening the tracker's cost matrix. The network itself is the same YOLO26 model from the official Ultralytics checkpoint. The raw callgrind files stay on the measuring machine and are not in git, so the committed ledger is a transcription.
Where the object detector on your home computer stops
A series page has to say where the claim ends. This one ends in three places.
An object detector on your home computer is host software
Every Diana number here was taken on a desktop-class x86-64 machine. No source in this release shows Diana on an ESP32, an ESP32-S3, or any firmware target. The boards dldeploy flashes run firmware built on Espressif's ESP-IDF, and their camera job is to serve frames, not to run YOLO. If detection ever joins a dldeploy setup, it belongs on the home computer, and this page treats it that way.
Instructions, not frames per second
34.6 percent fewer instructions does not mean 34.6 percent more frames per second. Callgrind does not see cache misses, page faults or thread barriers, and the campaign did not rerun the wall clock. Diana's older timing tables predate 0.10 and were not re-measured. So there is no new frame rate for an object detector on your home computer here, and this page will not invent one.
Two numeric defaults you can switch off
Two changes alter Diana's arithmetic slightly. Winograd convolution, from Lavin and Gray, runs on 3x3 layers with at least 16 channels, and attention softmax uses a polynomial exp. The repository reports a full-graph oracle pass on all five model sizes. FFAI_DIANA_NO_WINOGRAD=1 and FFAI_DIANA_SOFTMAX_LIBM=1 restore the old arithmetic. Tracking uses ByteTrack, from the ByteTrack paper, and its public API was kept.
How an object detector on your home computer could meet the camera
The series rule asks for two lists, then any proposal labeled as one.
What already runs on the LAN
The measured access is the camera page. A flashed ESP32 camera serves an MJPEG page that a browser on the same network opens directly, covered in the rusty_esp_video field note. Earlier measurements put it at about 13.8 to 14.3 frames per second under a 15 fps cap on a XIAO ESP32-S3 Sense. That is previous context, not a new benchmark. The comparison with buying a closed camera is a pure Rust alternative to a camera. The camera crates belong to the Janus family at remadewithrust.com, which is why that site is linked here.
What an object detector on your home computer still leaves unmeasured
- Diana reading frames from a dldeploy camera page: not built.
- Frames per second for any detector on a home computer: not measured in this campaign.
- The home computer's media controller receiving a Janus stream: not wired.
- Detection accuracy on real porch, nursery or hallway scenes: not measured.
- Reach from outside the house: not measured.
A proposed porch-camera flow, not a built one
One proposal: the home computer reads frames from the porch camera's MJPEG page on the LAN, decodes each JPEG, and hands the picture to Diana, so the picture is analyzed on your own machine and only a box and a label come out. That is a suggestion. No dldeploy page, tool or service does it today, and no frame rate for it has been measured. The machine that would run it is MATA's home computer, the subject of this series. Why a household would want detection at home at all is in a smart home without the cloud and when a baby monitor camera leaves the nursery. The camera vision is See.
FAQ
Quick answers for builders evaluating this technology.
Does Diana run on the ESP32 camera board?
No source in this release says so. The Diana measurements were taken on a desktop-class x86-64 machine under WSL. The ESP32 camera's job is to serve its MJPEG page on the LAN, and detection, if added later, would be host software on the home computer.
How many frames per second will Diana manage on my computer?
This campaign does not say. It counted instructions with callgrind on one image and did not rerun the clock. Diana's older timing tables predate version 0.10, so neither old nor new figures describe your machine.
What changed in Diana's tracker for home cameras?
The ByteTrack tracker's cost matrix became one flat array, and a tracker replay on cached MOT17-02 detections fell from 21,336,089 to 16,148,683 instructions, 24.3 percent. The public tracker API was kept.
Which camera access already runs on the LAN today?
The MJPEG page an ESP32 camera serves to a browser on the same network, measured earlier at about 13.8 to 14.3 frames per second under a 15 fps cap. Detection on those frames, the home computer's media controller and reach from outside are not built or measured.
Could Diana read frames from a porch camera page someday?
That is a proposal, not a feature. A home computer could decode frames from the camera's MJPEG page and hand them to Diana. No dldeploy page, tool or service does this today, and no frame rate for it has been measured.