Vision — SVP
The left panel of the Vision tab is shimeta-svp, a set of 9 built-in scene detection tools on the board that can be launched with one click, no training required.
Entry point
Top tab → "Vision" → the shimeta-svp panel on the left
The 9 scenes
| Button | Chinese name (UI) | Detection target |
|---|---|---|
| person | 人形检测 | Pedestrians |
| face | 人脸检测 | Faces |
| car | 车辆检测 | Motor vehicles |
| bike | 非机动车检测 | Bicycles, e-bikes |
| face_emo | 人脸情绪检测 | Smiling / neutral |
| dms | 驾驶员监测 | Closed eyes, yawning |
| head | 人头数检测 | Heads |
| fire | 火焰/烟雾检测 | Flames, smoke |
| person_kp | 人体关键点 | 17-point human skeleton |
Global parameters
The parameter panel at the top affects all scenes:
| Parameter | Default | Description |
|---|---|---|
| Confidence | 0.25 | Detections below this value are filtered out; lower = more detections |
| IOU | 0.45 | Higher = more suppression of duplicate boxes |
| H.264 | Checked | Encoding format of the RTSP stream |
After changing parameters, click "Run" to restart the service.
Starting detection
Steps
- Pick a scene button (e.g. "person")
- Click that scene card's "Run" button
- The shimeta-svp service starts (about 1–3 seconds)
- The "Object view" box in the upper right starts showing the live camera feed
Switching scenes
shimeta-svp runs only one scene at a time. To switch scenes:
- Click the current scene's "Run" button to stop it
- Pick the new scene
- Click the new scene's "Run" button
You cannot run multiple SVP scenes at the same time.
Camera feed (built into the Pico-G1)
The Pico-G1 board comes with an SC465SL camera (MIPI CSI interface).
Once shimeta-svp starts, the "Object view" box in the upper right shows the live camera feed directly:
- MJPEG stream (port 8080 /stream)
- Detection boxes overlaid on the image
- Class + confidence labels
The SC465SL is the only camera model officially supported by the Pico-G1; other cameras are not yet compatible.
Tuning parameters
If you get too many or too few detections:
Too few detections
Lower the confidence (0.25 → 0.1):
| Parameter | Default | More detections |
|---|---|---|
| Confidence | 0.25 | 0.1 |
Too many duplicate boxes
Raise the IOU (0.45 → 0.7):
| Parameter | Default | Suppresses duplicates |
|---|---|---|
| IOU | 0.45 | 0.7 |
After changing parameters you must restart the scene for them to take effect.
Performance reference
| Scene | Input size | Frame rate |
|---|---|---|
| person / car / bike | 640×640 | 22–24 fps |
| face / head | 640×640 | 22–24 fps |
| face_emo | 192×192 | 24 fps |
| dms | 320×240 | 22–24 fps |
| person_kp | 384×384 | 18–20 fps |
| fire | 640×640 | 22 fps |
Troubleshooting
| Symptom | Fix |
|---|---|
| Camera feed is black | Check the SC465SL ribbon cable connection |
| Clicking "Run" does nothing | shimeta-svp not installed; see Quick Start ch03 |
| Choppy feed | NPU temperature too high; improve the board's cooling |
| Port conflict with shimeta-xmm | The two cannot run at the same time; stop one first |
| All detection confidences below 5% | Model does not match the NPU; re-export via AIHub |
Next
- ch06 Vision — Custom Models: run deployed .xmm models
- ch09 Agent Debugging Assistant: let Claude Code tune the parameters for you
- ch08 Terminal: start/stop the SVP service manually