Can YOLO26-P2 Be Compiled to HEF on Hailo-8L

General Discussion
1

Hi everyone,

Has anyone successfully compiled a YOLO26-P2 model to a HEF for Hailo-8L?

I know YOLO26 support was added to the Hailo Model Zoo , but I couldn’t find any references specifically for the P2 variants with the additional high-resolution detection head.

I’m trying to determine:

  • whether YOLO26-P2 exports cleanly through ONNX → HAR → HEF,

  • if the current DFC supports the P2 architecture,

  • and whether any graph modifications or custom postprocessing are required.

If anyone has:

  • compiled YOLO26-P2 successfully,

  • encountered unsupported ops,

  • or has recommended export settings,

I’d appreciate any guidance.

Thanks!

2 Likes
2

As the extra head doesn’t include new layer types on top of those existing in the vanilla yolo26, no compilation issues are expected, albeit the FPS might come out a bit lower and similarly quantization performance may vary.

Regarding the postprocessing, you may find a generic solution in cutting the full ONNX into a HEF-translatable and a post-proc part, after figuring out the cutoff nodes using Netron. For an example script see this: hailo-apps/hailo_apps/python/core/common/extract_postprocessing.py at main · hailo-ai/hailo-apps

1 Like