Vision - Custom Models
The right panel of the Vision tab is shimeta-xmm, used to run custom .xmm models (from AIHub) deployed to the Pico-G1.
Entry point
Top tab → "Vision" → right-side shimeta-xmm panel
Panel states
Deployed models
If you have deployed a model in the Models tab before:
- The panel shows the list of deployed models
- You can select and run directly
No models yet
When you enter for the first time, the panel may show:
shimeta_xmm
暂无模型（部署到 /mnt/models/xmm/）This means the on-board /mnt/models/xmm/ directory is empty. See ch04 One-Click Deploy to deploy a model.
Parameter panel
| Parameter | Default | Description |
|---|---|---|
| conf | 0.25 | Confidence threshold |
| iou | 0.45 | NMS IoU threshold |
| Classes | All | Dropdown to select the classes to detect |
Class filtering
If the deployed model uses generic labels such as the COCO 80 classes, you can select from the dropdown:
| Option | Effect |
|---|---|
| All | Detect all classes |
| Single class | Detect only that class |
| Multi-select | Detect the selected classes |
Starting detection
- Confirm a model is deployed (the panel shows the model list)
- Set the conf / iou / Classes parameters
- Click the "Run" button
- The shimeta-xmm service starts (about 1-3 seconds)
- The "Objects" box in the upper-right corner shows the live camera view
Camera view
The Pico-G1 comes with a SC465SL camera (MIPI CSI interface).
After shimeta-xmm starts:
- MJPEG stream (port 8080 /stream)
- Detection boxes overlaid on the image (different colors per class)
- Class + confidence labels
shimeta-xmm and shimeta-svp share port 8080 — they cannot run at the same time.
Deploying a new model
To deploy a new model:
- Switch to the Models tab (ch03 Model Market)
- Select a model
- Click the "Deploy to board" button
- Return to the Vision tab when the deployment completes
- The new model appears automatically in the right-side xmm panel
Port conflicts
| Service | Ports |
|---|---|
| shimeta-svp | 8080 + 554 (RTSP) |
| shimeta-xmm | 8080 |
Both share 8080, so they cannot run at the same time. To switch:
- Stop the current service (click the running scene button)
- Start the new service
Performance reference
| Model | Input size | Frame rate |
|---|---|---|
| yolov8n | 640×640 | 15-22 fps |
| yolov5n | 640×640 | 18-25 fps |
| Small custom CNN | 320×320 | 25-30 fps |
Actual frame rate depends on model size and NPU utilization.
Troubleshooting
| Symptom | Fix |
|---|---|
| Panel shows "no models yet" | No model deployed; deploy one in the Models tab first |
| Clicking "Run" does nothing | Model file corrupted; redeploy |
| Camera view is black | Loose cable / incompatible camera (only SC465SL supported) |
| Port conflict | shimeta-svp is still running; stop it first |
| All detection confidences < 5% | Model does not match the NPU |
Next
- ch04 One-Click Deploy: deploy a new model
- ch05 Vision - SVP: 9 scene-based detections
- ch08 Terminal: manage on-board model files manually