CV / PP-YOLOE

PP-YOLOE

reComputer RK3588

PP-YOLOE COCO object detection accelerated by RKNN on reComputer RK3576 and RK3588.

27 downloads

Precision

RKNN

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

Platform

Size

Inference Engine

Getting Started

Armbian 26.8.7 Gnome or newer with OTA overlayroot: Configure Docker storage before running containers to avoid overlay mount errors.

Deploy

sudo docker run --rm --privileged --net=host --device /dev/dri/renderD128:/dev/dri/renderD128 \
  -v /proc/device-tree/compatible:/proc/device-tree/compatible:ro \
  ghcr.io/seeed-projects/recomputer-rk-cv/rk3588-ppyoloe:latest \
  python web_detection.py --platform rk3588 --model_path model/ppyoloe_s.rknn \
  --class_path model/coco_80_labels_list.txt --camera_id -1

REST API

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

Curl

curl http://localhost:8080/v1/chat/completions -d '{
  "model": "ppyoloe-rknn",
  "messages": [{"role": "user", "content": "Hello"}]
}'

Python

import requests

resp = requests.post(
    "http://localhost:8080/v1/chat/completions",
    json={"model": "ppyoloe-rknn", "messages": [{"role": "user", "content": "Hello"}]},
)
print(resp.json())

Model Details

Armbian 26.8.7 Gnome or newer with OTA overlayroot: Configure Docker storage before running containers to avoid overlay mount errors.

Quick Start

1. Install Docker

Run the following commands on the development board to install Docker:

bash

# Download installation script
curl -fsSL https://get.docker.com -o get-docker.sh
# Install using Aliyun mirror source
sudo sh get-docker.sh --mirror Aliyun
# Start Docker and enable auto-start on boot
sudo systemctl enable docker
sudo systemctl start docker

2. Run the Project (One command, dual-mode preview)

This deployment follows the PP-YOLOE modules in reComputer-RK-CV. The service provides RKNN NPU inference, Web preview, REST API, and asynchronous video analysis.

Model Information

Property	Value
Task	COCO 80-class object detection
Input	`640 x 640` letterboxed RGB
Variants	`ppyoloe_s.rknn`, `ppyoloe_m.rknn`
Default	`model/ppyoloe_s.rknn`
Post-processing	DFL decoding and class-aware NMS

Step A: Pull Images

bash

sudo docker pull ghcr.io/seeed-projects/recomputer-rk-cv/rk3588-ppyoloe:latest
sudo docker pull ghcr.io/seeed-projects/recomputer-rk-cv/rk3576-ppyoloe:latest

Step B: Run with One Click

For RK3576:

bash

sudo docker run --rm --privileged --net=host \
  --device /dev/dri/renderD128:/dev/dri/renderD128 \
  -v /proc/device-tree/compatible:/proc/device-tree/compatible:ro \
  ghcr.io/seeed-projects/recomputer-rk-cv/rk3576-ppyoloe:latest \
  python web_detection.py --platform rk3576 \
  --model_path model/ppyoloe_s.rknn \
  --class_path model/coco_80_labels_list.txt --camera_id -1

For RK3588:

bash

sudo docker run --rm --privileged --net=host \
  --device /dev/dri/renderD128:/dev/dri/renderD128 \
  -v /proc/device-tree/compatible:/proc/device-tree/compatible:ro \
  ghcr.io/seeed-projects/recomputer-rk-cv/rk3588-ppyoloe:latest \
  python web_detection.py --platform rk3588 \
  --model_path model/ppyoloe_s.rknn \
  --class_path model/coco_80_labels_list.txt --camera_id -1

Switch to PP-YOLOE-M with --model_path model/ppyoloe_m.rknn. For camera input, map /dev/videoN and set the matching --camera_id.

🔌 API Documentation

1. Model Inference Interface (Predict)

Endpoint: POST /api/models/ppyoloe/predict

Request Parameters (Multipart/Form-Data):

file: Optional image file to detect.
video: Optional MP4 file.
timestamp: Optional video timestamp in seconds; defaults to the first frame.
realtime: Optional boolean; with no upload, use the active camera or local-video frame.
conf: Optional request-level confidence threshold; defaults to 0.25.
iou: Optional request-level NMS IoU threshold; defaults to 0.45.

Usage Examples:

bash

curl -X POST "http://<BOARD_IP>:8000/api/models/ppyoloe/predict" \
  -F "file=@bus.jpg" -F "conf=0.30" -F "iou=0.45"

Response Format (JSON):

json

{
  "success": true,
  "source": "uploaded image",
  "predictions": [
    {
      "class": "bus",
      "confidence": 0.91,
      "box": {"x1": 120, "y1": 80, "x2": 520, "y2": 430}
    }
  ],
  "image": {"width": 640, "height": 480}
}

2. System Configuration Interface (Config)

Input priority is image, uploaded MP4 frame, then the active camera or sample video. GET/POST /api/config controls the global object and NMS thresholds. The runtime also exposes health, MJPEG, video upload, asynchronous analysis, status, list, and download interfaces.

web_detection.py performs letterbox preprocessing, RKNN inference, PP-YOLOE DFL decoding, class-aware NMS, coordinate restoration, rendering, and API serialization. Replacement models must remain compatible with that DFL decoder and output layout.

3. Command Line Arguments

Argument	Default	Description
`--platform`	Required	`rk3576` or `rk3588`.
`--model_path`	Required	PP-YOLOE-S or PP-YOLOE-M RKNN file.
`--class_path`	COCO labels	Class names used for serialization and drawing.
`--camera_id`	`1`	Camera index; `-1` enables upload-only mode.
`--video_path`	None	Looping local MP4; overrides the camera.
`--host` / `--port`	`0.0.0.0` / `8000`	Service address and port.

Real-time Video Stream Interface (Video Feed)

Get the latest annotated MJPEG stream for browser preview:

Endpoint: GET /api/video_feed
Example Usage: <img src="http://<BOARD_IP>:8000/api/video_feed">

🛠️ Developer Guide (Production Recommendations)

Build Local Images

bash

docker build -f docker/rk3576/ppyoloe.dockerfile \
  -t rk3576-ppyoloe:local src/rk3576_ppyoloe

docker build -f docker/rk3588/ppyoloe.dockerfile \
  -t rk3588-ppyoloe:local src/rk3588_ppyoloe

Inputs and Outputs

Input: image, video frame, or camera frame. Output: COCO classes, confidence scores, and boxes.