CV / NanoDet

NanoDet-RepVGG-a12

reComputer R Series (CM5 + Hailo-10H)

COCO 80-class object detection with nanodet_repvgg_a12 (YOLOX-based, on-chip NMS) on reComputer R Series with Hailo-8 or Hailo-10H.

40 downloads

Size

7.5 MB

Memory

4GB+

Precision

Hailo HEF / HailoRT

Choose the device you're using, the set up guide and documentation will update accordingly.

Platform

Inference Engine

Getting Started

Deploy

sudo docker run --rm \
  --name cm5-hailo10h-nanodet-repvgg-a12 \
  --privileged \
  --net=host \
  -e PYTHONUNBUFFERED=1 \
  --device /dev/hailo0:/dev/hailo0 \
  -v /usr/lib/libhailort.so.5.1.1:/usr/lib/libhailort.so.5.1.1:ro \
  -v /usr/lib/libhailort.so:/usr/lib/libhailort.so:ro \
  ghcr.io/seeed-projects/recomputer-hailo10h-cv/nanodet_repvgg_a12:latest \
  python web_detection.py --model_path model/nanodet_repvgg_a12.hef --video_path video/test.mp4

REST API

Use the REST API to run inference. Copy the commands below.

Curl

curl -X POST "http://<Board_IP>:8000/api/models/nanodet_repvgg_a12/predict" \
  -F "file=@test.jpg"

Python

import requests

response = requests.post(
    "http://<Board_IP>:8000/api/models/nanodet_repvgg_a12/predict",
    files={"file": open("test.jpg", "rb")},
    timeout=30,
)
print(response.json())

Model Details

reComputer R Series (CM5 + Hailo-10H)

NanoDet-RepVGG-a12 on reComputer CM5 + Hailo-10H

NanoDet-RepVGG performs COCO 80-class object detection on Hailo-10H with on-chip NMS (Hailo HPP). The HEF emits already-decoded detections and the app only parses the post-NMS tensor. Ported from the validated Hailo-8 module rpi5_hailo8_nanodet_repvgg_a12; the executor uses the HailoRT 5.1.1 create_infer_model multi-output wrapper with FLOAT32 outputs.

Model Info

Property	Value
Architecture	NanoDet-RepVGG (COCO-80)
Task	Object detection
Input	Read from the HEF at startup (printed on first run)
Output	on-chip NMS tensor, post-NMS shape 80x5x100
Classes	80 (COCO, 0-indexed — no labels_offset)
HEF	Hailo Model Zoo v5.4.0, Hailo-10H (4.6 MB)

Hardware and Host Setup

Item	Value
Board	reComputer R Series with Raspberry Pi CM5
Accelerator	Hailo-10H over PCIe, exposed as `/dev/hailo0`
Host driver	`hailo-h10-all` apt package
Runtime	HailoRT 5.1.1
Python in container	3.13, aarch64

bash

sudo apt update
sudo apt install hailo-h10-all -y
sudo reboot

# After reboot
hailortcli fw-control identify
ls /dev/hailo0

Run With Demo Video

bash

sudo docker run --rm \
  --name cm5-hailo10h-nanodet-repvgg-a12 \
  --privileged \
  --net=host \
  -e PYTHONUNBUFFERED=1 \
  --device /dev/hailo0:/dev/hailo0 \
  -v /usr/lib/libhailort.so.5.1.1:/usr/lib/libhailort.so.5.1.1:ro \
  -v /usr/lib/libhailort.so:/usr/lib/libhailort.so:ro \
  ghcr.io/seeed-projects/recomputer-hailo10h-cv/nanodet_repvgg_a12:latest \
  python web_detection.py --model_path model/nanodet_repvgg_a12.hef --video_path video/test.mp4

Open http://<Board_IP>:8000 to view the web preview. For a USB camera, mount /dev/video0 and replace --video_path video/test.mp4 with --camera_id 0.

REST API

bash

curl -X POST "http://<Board_IP>:8000/api/models/nanodet_repvgg_a12/predict" \
  -F "file=@test.jpg"

Endpoint	Method	Description
`/health`	GET	Health check
`/api/models/nanodet_repvgg_a12/predict`	POST	Detections (JSON)
`/api/video_feed`	GET	MJPEG preview stream

Implementation notes

The HEF runs NMS on-chip; the app only parses the post-NMS tensor, so nms_thresh is ignored (kept for API parity).
Post-NMS rows are [ymin, xmin, ymax, xmax, score], normalized to [0,1] of the letterboxed input; the app scales to pixels and un-letterboxes.
Input preprocessing (normalization / color conversion) is compiled into the HEF; the app letterboxes and feeds raw uint8 pixels.
HailoRT may return the NMS vstream as a ragged per-class list (NMS-by-score); the parser handles that plus object/dense layouts. First inference logs the raw type/shape for on-device verification.
Executor outputs are requested as FormatType.FLOAT32 (dequantized by the SDK); reading the raw quantized buffers yields values in the tens of thousands and zero valid detections.
Class mapping: cls_id (0..79) → standard COCO 80-class list directly.

Model source

nanodet_repvgg.hef is the Hailo-10H build from Hailo Model Zoo v5.4.0 — URL, SHA256 and size are recorded in the repository (src/hailo10h_nanodet_repvgg_a12/model/README.md).

Inputs and Outputs

Input: image, video, or USB camera frame. Output: COCO 80-class detection boxes with confidences, plus an annotated MJPEG preview.