CV / ResNet

ResNet50V2

ResNet50V2 image classification accelerated by RKNN on reComputer RK3576 and RK3588.

2 downloads

Precision

RKNN

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

Platform

Size

Inference Engine

Getting Started

Deploy

sudo docker run --rm --privileged --net=host \
  -e PYTHONUNBUFFERED=1 -e RKNN_LOG_LEVEL=0 \
  --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-resnet50v2:latest \
  python3 web_classification.py --model_path model/rk3576_resnet50-v2-7.rknn --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": "resnet50-v2-rknn",
  "messages": [{"role": "user", "content": "Hello"}]
}'

Python

import requests

resp = requests.post(
    "http://localhost:8080/v1/chat/completions",
    json={"model": "resnet50-v2-rknn", "messages": [{"role": "user", "content": "Hello"}]},
)
print(resp.json())

Model Details

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 page is adapted from the ResNet50V2 READMEs in reComputer-RK-CV. The service runs ImageNet classification on the Rockchip NPU and provides browser, MJPEG, health-check, and REST interfaces.

Model Information

Property	Value
Task	Image classification
Input	`224 x 224` RGB image
Output	Configurable Top-K ImageNet labels and confidence scores
RK3576 model	`model/rk3576_resnet50-v2-7.rknn`
RK3588 model	`model/rk3588_resnet50-v2-7.rknn`
Labels	`model/synset.txt`

Step A: Pull Images

bash

sudo docker pull ghcr.io/seeed-projects/recomputer-rk-cv/rk3588-resnet50v2:latest
sudo docker pull ghcr.io/seeed-projects/recomputer-rk-cv/rk3576-resnet50v2:latest

Step B: Run with One Click

For RK3576:

bash

sudo docker run --rm --privileged --net=host \
  -e PYTHONUNBUFFERED=1 -e RKNN_LOG_LEVEL=0 \
  --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-resnet50v2:latest \
  python3 web_classification.py \
  --model_path model/rk3576_resnet50-v2-7.rknn --camera_id -1

For RK3588:

bash

sudo docker run --rm --privileged --net=host \
  -e PYTHONUNBUFFERED=1 -e RKNN_LOG_LEVEL=0 \
  --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-resnet50v2:latest \
  python3 web_classification.py \
  --model_path model/rk3588_resnet50-v2-7.rknn --camera_id -1

Open http://<BOARD_IP>:8000 for the Web page or http://<BOARD_IP>:8000/docs for OpenAPI.

🔌 API Documentation

1. Model Inference Interface (Predict)

Endpoint: POST /api/models/resnet50v2/predict

Request Parameters (Multipart/Form-Data):

file: Optional image file to classify.
video: Optional MP4 file; it cannot be supplied together with file.
timestamp: Optional non-negative video timestamp in seconds; defaults to the first frame.
conf: Optional request-level confidence threshold from 0.0 to 1.0.
topk: Optional number of results from 1 to 20; defaults to 5.
With neither file nor video, the current camera or local-video frame is used when available.

Usage Examples:

bash

curl -X POST "http://<BOARD_IP>:8000/api/models/resnet50v2/predict" \
  -F "file=@dog_224x224.jpg" -F "topk=5" -F "conf=0.01"

bash

curl -X POST "http://<BOARD_IP>:8000/api/models/resnet50v2/predict" \
  -F "video=@test.mp4" -F "timestamp=5.5" -F "topk=3"

Response Format (JSON):

json

{
  "success": true,
  "model": "resnet50v2",
  "source": "uploaded image",
  "predictions": [
    {"class": "n02099601 golden retriever", "confidence": 0.89}
  ],
  "image": {"width": 224, "height": 224}
}

2. System Configuration Interface (Config)

GET /api/config returns conf_thresh; update it with JSON through POST /api/config. Request-level conf takes precedence.

Endpoint	Purpose
`GET /api/health`	Service, platform, model, and readiness.
`GET /api/video_feed`	Annotated MJPEG preview.
`POST /api/video/upload`	Upload an image or MP4.
`POST /api/video/analyze`	Analyze an uploaded `filename`.
`GET /api/video/status`	Processing progress and errors.
`GET /api/video/list`	Uploaded and generated files.
`GET /api/video/download/{filename}`	Download a result.

3. Command Line Arguments

The application accepts an uploaded image, a selected frame from an uploaded MP4, a looping local video, or a camera frame.

Argument	Default	Description
`--model_path`	Required	RKNN model selected for the target SoC.
`--camera_id`	`1`	Camera index; use `-1` for upload-only mode.
`--video_path`	None	Local MP4 path; it overrides the camera.
`--host`	`0.0.0.0`	FastAPI listen address.
`--port`	`8000`	Service port.

Map /dev/videoN into the container when selecting camera N.

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)

Model Replacement

The runtime resizes input to 224 x 224, converts BGR to RGB, applies softmax, and filters Top-K results. A replacement model must preserve the expected input layout, normalization, channel order, and output format. Update synset.txt if the class set changes. Runtime files are stored under workspace/; set RK_CV_WORKSPACE to override that location.

Build Local Images

bash

docker build -f docker/rk3576/resnet50v2.dockerfile \
  -t rk3576-resnet50v2:local src/rk3576_resnet50v2

docker build -f docker/rk3588/resnet50v2.dockerfile \
  -t rk3588-resnet50v2:local src/rk3588_resnet50v2

Inputs and Outputs

Input: image, video frame, or camera frame. Output: Top-K ImageNet labels and confidence scores.