SVP Scene Detection
shimeta_svp includes multiple pre-compiled GNN models covering 9 scenario types. No model file needed — works out of the box.
First-time Setup
Download the model pack (62MB, one-time operation):
. /mnt/setup.sh
wget https://mirrors.shimetapi.cn/repos/GK7206/install_models.sh -O - | shQuick Commands
# Person / Vehicle basic detection
shimeta_svp person # Person detection
shimeta_svp face # Face detection
shimeta_svp car # Vehicle detection (motor vehicles)
shimeta_svp bike # Non-motor vehicle detection (bicycle/e-bike)
# Advanced scenarios
shimeta_svp face_emo # Facial emotion detection (smile/neutral)
shimeta_svp dms # Driver drowsiness monitoring (eye closure/yawning)
shimeta_svp head # Head count detection
shimeta_svp fire # Fire/smoke detection
shimeta_svp person_kp # 17-point human keypointsViewing Results
After starting, an RTSP video stream is generated at rtsp://192.168.49.10:554/livestream/0 with real-time detection overlays (hardware-drawn boxes, zero CPU overhead).
How to play the RTSP stream:
- Install ffmpeg (tutorial: https://cloud.tencent.com/developer/article/2470274)
- Press
Win+R, typecmdto open the command line, then run:
ffplay rtsp://192.168.49.10:554/livestream/0Parameter Tuning
shimeta_svp person 0.1 # Lower threshold — detect more
shimeta_svp car 0.5 0.3 # Higher threshold + fewer duplicates
shimeta_svp face 0.25 0.45 h264 # H264 encoding (better VLC/ffplay compatibility)Default is H265 encoding. If VLC/ffplay shows corrupted frames or
Skipping invalid NALU, add theh264parameter to switch to H264 encoding.
SVP vs shimeta_camera
shimeta_svp | shimeta_camera | |
|---|---|---|
| Models | Multiple built-in pre-compiled models | Requires uploading .xmm.zip |
| Ease of use | One command, no training needed | Requires model exported from training platform |
| Frame rate | 22-24fps | 15-16fps |
| Output | RTSP stream (VLC/ffplay) | MJPEG WebUI (browser) |
| Best for | Quick demos / competitions | Custom model tuning |