UPDATE - SOLVED ✓
Problem Summary:
- YOLOv8-seg custom model (tire detection, 4 classes)
- GStreamerInstanceSegmentationApp only supports YOLOv5-seg
- Missing post-processing library for YOLOv8-seg
Solution: DeGirum PySDK
The Hailo community guide pointed me to DeGirum PySDK, which has native YOLOv8-seg support.
What I did:
- HEF compilation (hailomz) - ✓ No issues
- Install DeGirum:
pip install degirum - Create Model Zoo directory with:
best.hef(compiled model)config.json(model configuration + Checksum required!)labels.json(class names)
- Run inference with simple Python code
Results:
- ✓ Real-time webcam: ~10 FPS (Raspberry Pi 5 + Hailo-8)
Key Points:
- DeGirum handles YOLOv8-seg post-processing internally
- No need to write custom .so files
- JSON configuration is straightforward
- Webcam requires GStreamer backend (not V4L2)
Resources:
- DeGirum Guide: User Guide 4: Simplifying Instance Segmentation on a Hailo Device Using DeGirum PySDK
Big thanks to the Hailo team and community for the excellent support!
I’ll be back if I run into any new issues during development! ![]()