Sorter V2
Detector Documentation Hub
This site is the durable documentation layer for detector benchmarking, model artifacts, and target conversion workflows across Mac, Raspberry Pi, Orange Pi, and Hailo.
Why this site exists
The detector work produced a lot of useful results, but also a lot of local experiment output:
- benchmark JSONs
- HTML comparison reports
- compile bundles
- Vast.ai session artifacts
The goal of this site is to keep the lasting knowledge in one checked-in place and treat generated output as local working state unless it is explicitly promoted.
Documentation map
The current conclusions, deployment recommendations, and the canonical local artifacts worth keeping.
What each export and compiled format is for, where it lives, and which target it serves.
The repeatable cross-device benchmark workflow for Mac, Orange Pi, and Raspberry Pi.
The maintained ONNX to HEF flow for Raspberry Pi 5 AI HAT deployments.
Current headline conclusions
- The local Mac Mini M4 CPU run is the current quality reference.
- Mac CoreML is the fastest validated local acceleration path, especially for
YOLO11s. - Orange Pi CPU and Raspberry Pi 5 CPU match the reference exactly on the shared benchmark bundle.
- Raspberry Pi 5
Hailo + NanoDetis the strongest current accelerated deployment path. - The current Orange Pi
RKNNartifacts are still experimental because they were not rebuilt from the exact current ONNX exports.
Artifact policy
The durable rule is simple:
- Keep decisions and workflows in checked-in Markdown.
- Keep only the latest canonical benchmark and compile artifacts under
software/client/blob/. - Regenerate reports from benchmark JSONs instead of treating every generated HTML file as permanent.
GitHub Pages publishing
This site is intended to publish from the root-level docs/ folder via GitHub Actions.
If the repository’s default branch is not main or master, update the Pages workflow trigger in .github/workflows/documentation-pages.yml after merge.