shimeta_camera — Real-time Camera Inference

If the board has a connected SC465SL camera, you can load your custom model for real-time inference.

Quick Start

. /mnt/setup.sh
shimeta_camera /mnt/yolov8n_demo.zip

Open http://192.168.49.10:8080 in a browser to view:

MJPEG live video (hardware VENC encoded, smooth 15-30fps)
Detection box overlay (frontend Canvas rendered, color-coded by class)
Detection results table (class, confidence, coordinates)
Real-time status panel (total frames, FPS, inference latency)

Parameter Tuning

shimeta_camera /mnt/yolov8n_demo.zip 0.1        # Lower threshold
shimeta_camera /mnt/yolov8n_demo.zip 0.25 0.3   # Reduce duplicate boxes

Note: Requires camera hardware. If the board doesn't have a camera, use shimeta_infer for image inference.

HTTP API Endpoints

When shimeta_camera is running, it exposes HTTP services on port 8080:

Endpoint	Method	Description
`/`	GET	WebUI dashboard (HTML)
`/stream`	GET	MJPEG live video stream
`/api/status`	GET	Detection results as JSON
`/api/snapshot`	GET	Current frame JPEG snapshot

`/api/status` JSON Format

{
  "total_frames": 150,
  "fps": 15.2,
  "infer_ms": 35.7,
  "detections": [
    {
      "class_id": 0,
      "label": "person",
      "conf": 0.85,
      "x1": 225,
      "y1": 313,
      "x2": 393,
      "y2": 638
    },
    {
      "class_id": 2,
      "label": "car",
      "conf": 0.73,
      "x1": 500,
      "y1": 200,
      "x2": 600,
      "y2": 400
    }
  ]
}

Key Performance Specs

Parameter	Value
NPU inference	~35ms (YOLOv8n 640×640)
Y→RGB NEON copy	~5ms
Post-process INT8 decode	~5ms
Actual frame rate	15-16fps (VPSS 3-channel pipeline)
Model input	640×640 uint8