Artifact Registry

Model Artifacts

Which detector artifacts currently exist, what each format is for, and how each target path is produced from training outputs.

Scope

The current maintained detector set is:

20260331-zone-classification-chamber-yolo11s
20260331-zone-classification-chamber-nanodet

These are the chamber-zone detector models used for the April 6, 2026 cross-device benchmark work.

Canonical source runs

YOLO11s

Source run:
- software/client/blob/local_detection_models/20260331-zone-classification_chamber-yolo11s/run.json
Canonical ONNX export:
- software/client/blob/local_detection_models/20260331-zone-classification_chamber-yolo11s/exports/best.onnx
Canonical benchmark bundle copy:
- software/client/blob/device_benchmarks/chamber_zone_pair_bundle/models/20260331-zone-classification-chamber-yolo11s/model.onnx

NanoDet

Source run:
- software/client/blob/local_detection_models/20260331-zone-classification_chamber-nanodet/run.json
Canonical ONNX export:
- software/client/blob/local_detection_models/20260331-zone-classification_chamber-nanodet/exports/best.onnx
Canonical benchmark bundle copy:
- software/client/blob/device_benchmarks/chamber_zone_pair_bundle/models/20260331-zone-classification-chamber-nanodet/model.onnx

Format matrix

Format	Purpose	Current status	Canonical location
`ONNX`	Main interchange format and correctness reference on CPU paths	Approved	`software/client/blob/device_benchmarks/chamber_zone_pair_bundle/models/.../model.onnx`
`NCNN`	CPU deployment experiments on low-power devices	Built, but currently not quality-approved for these chamber-zone exports	`software/client/blob/device_benchmarks/chamber_zone_pair_bundle/models/.../model.ncnn.param` and `model.ncnn.bin`
`CoreMLExecutionProvider`	Fast local Mac acceleration path using ONNX Runtime	Approved for local benchmarking; no separate `.mlpackage` is maintained right now	Reuses the canonical ONNX models
`RKNN`	Orange Pi NPU deployment path	Experimental; current artifacts are not rebuilt from the exact current ONNX exports	Existing Orange Pi device-side artifacts under `/root/bench/models`
`HEF`	Raspberry Pi 5 AI HAT deployment path	Built for both models; NanoDet currently has the stronger quality result	`software/client/blob/hailo_compile_bundles/.../results/`
`HAR`	Intermediate Hailo compiler output	Built and worth keeping with the matching `HEF`	`software/client/blob/hailo_compile_bundles/.../results/`

Current target mapping

Mac Mini M4

CPU reference:

Input format: ONNX
Runtime: onnxruntime on CPU
Status: reference path

Accelerated local path:

Input format: ONNX
Runtime: onnxruntime with CoreMLExecutionProvider
Status: benchmark-approved

Raspberry Pi 5

CPU fallback and correctness path:

Input format: ONNX
Runtime: onnxruntime==1.23.2
Status: approved

AI HAT path:

Input format: HEF
Runtime: Hailo runtime stack on Raspberry Pi OS
Status:
- NanoDet: strongest current accelerated result
- YOLO11s: built and runnable, but still needs closer parity tuning

Orange Pi 5

CPU fallback and correctness path:

Input format: ONNX
Runtime: onnxruntime
Status: approved

NPU path:

Input format: RKNN
Runtime: rknnlite
Status: experimental until rebuilt from the exact current ONNX exports

Conversion rule

The durable rule is:

training run -> canonical ONNX export -> target-specific compiled format

Do not compile target-specific formats directly from ad hoc copies if the ONNX source is unclear.

Conversion map

1. Training run to ONNX

The training and export runs live under:

software/client/blob/local_detection_models/...

The ONNX export from that run is the canonical source for downstream targets.

2. ONNX to benchmark bundle

Use:

cd software/client
uv run python scripts/device_detector_benchmark.py bundle \
  --preset chamber_zone_pair \
  --output blob/device_benchmarks/chamber_zone_pair_bundle \
  --archive

The bundle becomes the single benchmarking input across devices.

3. ONNX to Mac CPU, Mac CoreML, Pi CPU, Orange Pi CPU

Use:

uv run python scripts/device_detector_benchmark.py run \
  --bundle blob/device_benchmarks/chamber_zone_pair_bundle \
  --output-dir <target-output-dir> \
  --runtime onnx

For the Mac accelerated path:

uv run python scripts/device_detector_benchmark.py run \
  --bundle blob/device_benchmarks/chamber_zone_pair_bundle \
  --output-dir <target-output-dir> \
  --runtime coreml

4. ONNX to NCNN

The benchmark bundle already carries NCNN exports, but the current chamber-zone NCNN results are not approved yet.

5. ONNX to RKNN

The Orange Pi NPU path needs:

the exact current ONNX export
a matching RKNN toolkit and compiler environment
a fresh .rknn rebuild

Current status: runtime-side RKNN execution exists, but the current chamber-zone RKNN artifacts are not yet quality-approved.

6. ONNX to Hailo HEF

Use:

software/client/scripts/prepare_hailo_compile_bundle.py
software/client/scripts/vastai_hailo_session.py

Detailed instructions live in Hailo HEF Workflow.

Current maintained results:

software/client/blob/hailo_compile_bundles/classification_chamber_yolo11s/results/yolov11s_piece_320.hef
software/client/blob/hailo_compile_bundles/classification_chamber_nanodet/results/nanodet_plus_m_1_5x_piece_416_raw.hef

Update policy

When a new detector export replaces the current chamber-zone models, update:

this page
Runtime Status
Device Benchmarking if the benchmark preset or runtime rules change
Hailo HEF Workflow if the Hailo compile flow changes